fix(qwen3_moe): export W4A16 NVFP4 checkpoints via NVFP4 MoE path

[potoior]

Summary

QUANT_NVFP4_A16 was introduced in 0.10.0 to distinguish W4A16 weight-only
NVFP4 from W4A4 NVFP4, but Qwen3SparseMoeBlock._prepare_moe_weights only
gated on QUANT_NVFP4. Mixed-precision W4A16_NVFP4 checkpoints such as
Qwen3.6-35B-A3B-NVFP4 therefore fell through to the INT4 GPTQ Marlin
repack
and crashed with:

  AttributeError: 'NVFP4A16MarlinLinear' object has no attribute
  'qweight'

Changes

• modeling_qwen3_moe.py: treat QUANT_NVFP4_A16 as NVFP4 when gating the
  NVFP4 MoE plugin path.
• repacking.py: accept NVFP4A16MarlinLinear experts in
  repack_nvfp4_gated_moe_experts (their raw weight/weight_scale/
  weight_scale_2 buffers are the same ModelOpt NVFP4 format, so
  decode_modelopt_nvfp4 reuses them as-is).

Type of change

Bug fix

Verification

tensorrt-edgellm-export Qwen3.6-35B-A3B-NVFP4 --skip-audio --skip-visual
completes successfully after the fix (40-layer MoE repacked into
Nvfp4MoePlugin).

Compatibility

Backward compatible

Checklist

• I have run the hooks manually with pre-commit run --all-files and
  fixed any reported issues.
• All tests are passing.
• The change is backward compatible

[nvluxiaoz]

Thanks for the contribution! We will formalize W4A4 vs W4A16 support in following releases.

[potoior]

@nvluxiaoz Friendly ping — could you verify this fix directly? Once verified, this PR should be ready to merge. Thanks!

[sylvesterkaczmarek]

Could we add a focused regression test for the QUANT_NVFP4_A16 MoE path?

This fix depends on two separate conditions staying aligned: Qwen3 must select the NVFP4 MoE path for QUANT_NVFP4_A16, and repack_nvfp4_gated_moe_experts must accept NVFP4A16MarlinLinear. A small test exercising those together would make it much harder for a future quant-type refactor to send W4A16 experts back through the INT4 path.

The successful 35B export is useful validation, but it would be good to have this behaviour pinned in CI as well.