[Inference API] Support multimodal inputs for chat completion (#142736)

This commit allows "image_url" and "file" to be specified as types of
"content" in chat completion requests for ElasticInferenceService. The
specification matches the OpenAI specification for multimodal chat
completion inputs.

Other services will throw an UnsupportedOperationException if they
receive a chat completion request with multimodal inputs.

Content objects with the type "image_url" are specified as:
{
  "type": "image_url",
  "image_url": {
    "url": "base64 encoded image data or URL (if supported)",
    "detail": "optional detail value"
  }
}

Content objects with the type "file" are specified as:
{
  "type": "file",
  "file": {
    "file_data": "base64 encoded file data",
    "filename": "file name"
  }
}
Support for the "file_id" field in the "file" object is not added in
this commit.

When sending the chat completion request to the Elastic Inference
Service, the same OpenAI-compatible schema is used as when sending the
request to the inference API, so no additional translation logic between
the two is needed.

Other changes in this commit:
- Convert ContentObject class to an abstract class which is extended by
  ContentObjectText, ContentObjectImage and ContentObjectFile
- Modify serialization for ContentObject to throw an exception if
  attempting to send non-text content to an older version node
- Add backward compatibility tests for the above serialization change
- Require that "type" is one of "text", "image_url" or "file" for
  "content" objects. Previously any arbitrary value was allowed.
- Update existing tests to use ContentObjectText
- Add tests for default behaviour of embedding task and unified chat
  completion with multimodal inputs to SenderServiceTests

Author: DonalEvans

docs/changelog/142736.yaml

@@ -0,0 +1,5 @@
+area: Inference
+issues: []
+pr: 142736
+summary: "[Inference API] Support multimodal inputs for chat completion"
+type: enhancement

server/src/main/java/org/elasticsearch/inference/UnifiedCompletionRequest.java

@@ -0,0 +1 @@
+9290000

server/src/main/resources/transport/definitions/referable/inference_api_multimodal_chat_completion.csv

@@ -1 +1 @@
-esql_topn_min_competitive_updates,9289000
+inference_api_multimodal_chat_completion,9290000

server/src/main/resources/transport/upper_bounds/9.4.csv

@@ -7,18 +7,24 @@
 
 package org.elasticsearch.xpack.core.inference.action;
 
+import org.elasticsearch.ElasticsearchStatusException;
 import org.elasticsearch.TransportVersion;
 import org.elasticsearch.common.io.stream.NamedWriteableRegistry;
 import org.elasticsearch.common.io.stream.Writeable;
 import org.elasticsearch.core.TimeValue;
 import org.elasticsearch.inference.TaskType;
 import org.elasticsearch.inference.UnifiedCompletionRequest;
+import org.elasticsearch.rest.RestStatus;
 import org.elasticsearch.xpack.core.inference.InferenceContext;
 import org.elasticsearch.xpack.core.ml.AbstractBWCWireSerializationTestCase;
 
 import java.io.IOException;
+import java.util.Collection;
 import java.util.List;
+import java.util.function.Predicate;
 
+import static org.elasticsearch.inference.UnifiedCompletionRequest.MULTIMODAL_CHAT_COMPLETION_SUPPORT_ADDED;
+import static org.elasticsearch.test.BWCVersions.DEFAULT_BWC_VERSIONS;
 import static org.hamcrest.Matchers.is;
 
 public class UnifiedCompletionActionRequestTests extends AbstractBWCWireSerializationTestCase<UnifiedCompletionAction.Request> {
@@ -70,17 +76,53 @@ public void testValidation_ReturnsNull_When_TaskType_IsAny() {
 
     @Override
     protected UnifiedCompletionAction.Request mutateInstanceForVersion(UnifiedCompletionAction.Request instance, TransportVersion version) {
+        InferenceContext context = instance.getContext();
         if (version.supports(INFERENCE_CONTEXT) == false) {
-            return new UnifiedCompletionAction.Request(
-                instance.getInferenceEntityId(),
-                instance.getTaskType(),
-                instance.getUnifiedCompletionRequest(),
-                InferenceContext.EMPTY_INSTANCE,
-                instance.getTimeout()
-            );
+            context = InferenceContext.EMPTY_INSTANCE;
         }
 
-        return instance;
+        return new UnifiedCompletionAction.Request(
+            instance.getInferenceEntityId(),
+            instance.getTaskType(),
+            instance.getUnifiedCompletionRequest(),
+            context,
+            instance.getTimeout()
+        );
+    }
+
+    // Versions before MULTIMODAL_CHAT_COMPLETION_SUPPORT_ADDED throw an exception when serializing non-text content
+    // Those are tested in testMultimodalContentIsNotBackwardsCompatible
+    @Override
+    protected Collection<TransportVersion> bwcVersions() {
+        return super.bwcVersions().stream().filter(version -> version.supports(MULTIMODAL_CHAT_COMPLETION_SUPPORT_ADDED)).toList();
+    }
+
+    public void testMultimodalContentIsNotBackwardsCompatible() throws IOException {
+        var unsupportedVersions = DEFAULT_BWC_VERSIONS.stream()
+            .filter(Predicate.not(version -> version.supports(MULTIMODAL_CHAT_COMPLETION_SUPPORT_ADDED)))
+            .toList();
+        for (int runs = 0; runs < NUMBER_OF_TEST_RUNS; runs++) {
+            var testInstance = createTestInstance();
+            for (var unsupportedVersion : unsupportedVersions) {
+                if (testInstance.getUnifiedCompletionRequest().containsMultimodalContent()) {
+                    var statusException = assertThrows(
+                        ElasticsearchStatusException.class,
+                        () -> copyWriteable(testInstance, getNamedWriteableRegistry(), instanceReader(), unsupportedVersion)
+                    );
+                    assertThat(statusException.status(), is(RestStatus.BAD_REQUEST));
+                    assertThat(
+                        statusException.getMessage(),
+                        is(
+                            "Cannot send a multimodal chat completion request to an older node. "
+                                + "Please wait until all nodes are upgraded before using multimodal chat completion inputs"
+                        )
+                    );
+                } else {
+                    // If the instance doesn't contain multimodal content, assert that it can still be serialized
+                    assertBwcSerialization(testInstance, unsupportedVersion);
+                }
+            }
+        }
     }
 
     @Override