Add vision-only Edge-LLM ExecuTorch delegate integration

[micwill755]

Summary

Adds a vision-only TensorRT-Edge-LLM delegate path for ExecuTorch.

• Adds edge_llm::vision_tower and EdgeLLMPartitioner.
• Adds EL01 metadata wrapping existing TR01/TR02 engine payloads.
• Adds native EdgeLLMBackend and payload parsing.
• Adds PI0.5 HWC vision export and .pte packaging.
• Extends the reference runner for FP16 inputs.
• Preserves the existing generic TensorRTBackend path.

Motivation and context

PI0.5 components require Edge-LLM-specific runtime behavior beyond generic TensorRT engine execution. This change establishes the first component path—vision—while keeping the exported program libtorch-free.

The partitioner maps edge_llm::vision_tower to EdgeLLMBackend. At runtime, the backend creates a stateful Edge vision runner and invokes it through ExecuTorch’s delegate interface.

A fresh PI0.5 vision engine was exported and executed successfully:

input:  [4, 224, 224, 3] FP16
output: [1024, 2048] FP16

Dependencies

• ExecuTorch 1.4.1 or the repository’s pinned ExecuTorch commit.
• CUDA 13 and TensorRT 11.2.
• TensorRT-Edge-LLM companion changes providing libedgellmExecutorch.so.
• libNvInfer_edgellm_plugin.so available through EDGELLM_PLUGIN_PATH.
• Edge support is optional and feature-gated; normal Torch-TensorRT builds remain unchanged.

Issue

No tracking issue was provided.

Type of change

• New feature (non-breaking change which adds functionality)

Testing

• 7 focused Python tests passed.
• C++ EL01 parser Bazel test passed.
• Edge runtime and native backend built successfully.
• Fresh PI0.5 vision .pte executed successfully.
• Existing TensorRT .pte runner regression passed.

Checklist

• My code follows the style guidelines of this project.
• I have performed a self-review of my own code.
• I have commented hard-to-understand areas and integration constraints.
• I have made corresponding documentation changes.
• I have added tests to verify the feature.
• New and existing relevant unit tests pass locally.
• I have added the relevant PR labels and notified reviewers.