Update on "[Executorch][target recipes] Add target based recipes for lowering models to a target device"

This diff introduces multi backend/ target based recipes to lower a model with very less code. Target recipes provide pre-configured backend recipes to use them and retarget if needed.

See RFC: #13732

## Usage

```
from executorch.export import export
from executorch.export.target_recipes import get_ios_recipe

# CoreML + XNNPACK (FP32)
recipe = get_ios_recipe() # default = "ios-arm64-coreml-fp16"
session = export(model, recipe, example_inputs)
session.save_pte_file("model.pte")
```

## Advanced usage to combine specific recipes of one or two or backends
one can directly use `ExportRecipe.combine_recipes()` to combine specific backend recipes.
```
recipe1 = ExportRecipe.get_recipe(CoreMLRecipeType.FP32)
recipe2 = ExportRecipe.get_recipe(XNNPackRecipeType.FP32)
combined_recipe = ExportRecipe.combine(
                     [recipe1, recipe2],
                     recipe_name="multi_backend_coreml_xnnpack_fp32"
                  )
session = export(model, combined_recipe, example_inputs)
```
Additional changes:
1. Relaxed kwarg validation in the backend providers to just ignore them instead of erroring out.

Fixes: #13732

Differential Revision: [D81297451](https://our.internmc.facebook.com/intern/diff/D81297451/)

[ghstack-poisoned]

Author: abhinaykukkadapu

backends/xnnpack/recipes/xnnpack_recipe_types.py

@@ -12,23 +12,25 @@
 class XNNPackRecipeType(RecipeType):
     """XNNPACK-specific recipe types"""
 
-    FP32 = "fp32"
+    FP32 = "xnnpack_fp32"
 
     ## PT2E-based quantization recipes
     # INT8 Dynamic Quantization
-    PT2E_INT8_DYNAMIC_PER_CHANNEL = "pt2e_int8_dynamic_per_channel"
+    PT2E_INT8_DYNAMIC_PER_CHANNEL = "xnnpack_pt2e_int8_dynamic_per_channel"
     # INT8 Static Quantization, needs calibration dataset
-    PT2E_INT8_STATIC_PER_CHANNEL = "pt2e_int8_static_per_channel"
-    PT2E_INT8_STATIC_PER_TENSOR = "pt2e_int8_static_per_tensor"
+    PT2E_INT8_STATIC_PER_CHANNEL = "xnnpack_pt2e_int8_static_per_channel"
+    PT2E_INT8_STATIC_PER_TENSOR = "xnnpack_pt2e_int8_static_per_tensor"
 
     ## TorchAO-based quantization recipes
     # INT8 Dynamic Activations INT4 Weight Quantization, Axis = 0
     TORCHAO_INT8_DYNAMIC_ACT_INT4_WEIGHT_PER_CHANNEL = (
-        "torchao_int8da_int4w_per_channel"
+        "xnnpack_torchao_int8da_int4w_per_channel"
     )
     # INT8 Dynamic Activations INT4 Weight Quantization, default group_size = 32
     # can be overriden by group_size kwarg
-    TORCHAO_INT8_DYNAMIC_ACT_INT4_WEIGHT_PER_TENSOR = "torchao_int8da_int4w_per_tensor"
+    TORCHAO_INT8_DYNAMIC_ACT_INT4_WEIGHT_PER_TENSOR = (
+        "xnnpack_torchao_int8da_int4w_per_tensor"
+    )
 
     @classmethod
     def get_backend_name(cls) -> str:

export/recipe.py

@@ -301,6 +301,9 @@ def _combine_recipes(  # noqa: C901
                 edge_compile_config=edge_compile_config or EdgeCompileConfig(),
             )
 
+        recipe_name = recipe_name or "_".join(
+            [r.name for r in backend_recipes if r.name is not None]
+        )
         return cls(
             name=recipe_name,
             quantization_recipe=combined_quantization_recipe,