Alpamayo Edge exporter support

[micwill755]

Summary

• add Alpamayo 1.5 as a unified Edge exporter model family
• restore base or ModelOpt-quantized checkpoints and compile the original Qwen3-VL vision/language and Alpamayo action modules
• use temporary family setattr patches for fixed-grid vision, plugin-backed language prefill, and explicit stacked-prefix-KV diffusion steps
• keep helpers limited to multimodal tensor preparation and fixed-grid state setup
• add unified runner arguments and focused registration/packing/action coverage

Validation

• pre-commit hooks pass for every changed file
• Python syntax compilation passes
• Alpamayo-focused tests: 4 passed
• full exporter file: 22 passed, 1 pre-existing GR00T failure (apply_groot_patches is absent on the parent branch)
• Edge-LLM TensorRT plugin builds and loads successfully against CUDA 13 / TensorRT 11
• full GPU engine compilation remains to be run

Stack

This PR is stacked on #4723 (exporters-unified-runner).