feat: add completion types and optimize union unmarshal (envoyproxy#1242)

**Description**

This PR adds OpenAI's legacy `/completions` endpoint support to Envoy AI
Gateway, enabling compatibility with models like `babbage-002` and
`gpt-3.5-turbo-instruct`. Ideal for cost-optimized or fine-tuned setups
(e.g., vLLM/llama.cpp) needing suffix params or token prompts.

Refactored unions (`ContentUnion`, `EmbeddingRequestInput`,
`PromptUnion`) with custom unmarshaling for **1.6x–10x speedups**,
**60%–95% less memory**, and **52%–93% fewer allocations** vs.
openai-go. Gains are critical for batch/token-heavy workloads.

*Key Changes*
- New types: `CompletionRequest`, `CompletionResponse`, `PromptUnion`.
- Model constants for legacy testing.
- Unified `Usage` across APIs.
- New Benchmarks!

**Related Issues/PRs**

Starts envoyproxy#1231 

---------

Signed-off-by: Adrian Cole <[email protected]>

Author: codefromthecrypt

.codespell.skip

@@ -11,3 +11,5 @@ go.sum
 ./tests/e2e/logs
 *_for_tests.yaml
 ./tests/extproc/testdata/server.*
+./tests/internal/testopenai/cassettes/*.yaml
+./tests/internal/testopeninference/spans/*.json

cmd/aigw/docker-compose.yaml

@@ -77,6 +77,6 @@ services:
           -X POST http://aigw:1975/v1/embeddings \
           -H "Authorization: Bearer unused" \
           -H "Content-Type: application/json" \
-          -d "{\"model\":\"$$EMBEDDINGS_MODEL\",\"input\":\"What is RAG?\"}"
+          -d "{\"model\":\"$$EMBEDDINGS_MODEL\",\"input\":\"How do I reset my password?\"}"
     extra_hosts:  # localhost:host-gateway trick doesn't work with aigw
       - "host.docker.internal:host-gateway"