[AutoDeploy] Standardized MoE handling + onboarding

MoE Interface discussion

• 3 stages as shown below
• Each stage should be one or a set of self-contained transforms
• Separate each stage into separate file in the transform library

Stage 1 (mapping to unquantized reference op)

  # THIS STAGE: EVERYTHING GOES TO torch_moe op
  match_moe_pattern:
    # TODO: deprecate this one since not used anymore
    stage: pattern_matcher
  match_dense_moe_pattern:
    # TODO: make this one match to torch_moe instead of torch_moe_dense_mlp (used by GPT-OSS)
    stage: pattern_matcher
  match_bmm_moe_pattern:
    # TODO: put functionality into match_dense_moe_pattern and deprecate this one (used by Llama4)
    # TODO: understand if load hook functionality is needed and/or can be integrated into match_dense_moe_pattern
    stage: pattern_matcher

Stage 2 (mapping to quantized reference op)

# THIS STAGE: GOES FROM torch_moe TO torch_quant_*_moe op (torch_quant_fp8_moe, torch_quant_nvfp4_moe, torch_quant_mxfp4_moe)
  quantize_fp8_moe:
    stage: pattern_matcher
  quantize_nvfp4_moe:
    stage: pattern_matcher
  quantize_mxfp4_moe:
    stage: pattern_matcher

Stage 3 (mapping to optimal, fused kernels)

  # THIS STAGE: map to optimal kernel backend for MoE (triton, trtllm, cutlass, etc...). from torch_**_moe to {backend}_**_moe
  fuse_moe:
    stage: post_load_fusion
    enabled: true
    backend: trtllm
  fuse_fp8_moe:
    stage: post_load_fusion
    enabled: true
    backend: trtllm
  fuse_nvfp4_moe:
    stage: post_load_fusion
    enabled: false
  # TODO: Implement this one, doesn't exist yet
  fuse_mxfp4_moe:
    stage: post_load_fusion
    enabled: false
    backend: triton