Fix Qwen3.5 MTP sanitize norm shift

[xxxkkw]

Summary

• Stop using stripped mtp.* weights as a signal to shift Qwen3.5 norm weights by +1 during sanitize.
• Keep the raw-checkpoint norm shift tied to unsanitized Conv1d layout conversion.
• Cover both qwen3_5 and qwen3_5_moe wrapper paths so already-converted checkpoints with MTP weights can be loaded without shifting norms twice.

Environment

• OS: macOS Darwin 25.4.0 arm64
• Hardware: Apple M1 Max, 32 GB unified memory
• Python: local project virtual environment

Testing

• python -m unittest discover -s tests -p test_models.py -k qwen3_5_family_convert_then_load_norm_not_shift_twice
  • Result: 1 passed

Benchmark / profiling notes

• This is a sanitize/load correctness fix and does not change inference kernels.
• The regression test reproduces the load-sanitize path with already-converted norm weights plus stripped MTP weights; before this change the MTP key alone triggered an extra norm shift.

Context

• Full native MTP decoding is already being discussed in Native MTP speculative decoding (Qwen3.5/3.6 reference implementation) #990; this PR only extracts the small sanitize correctness fix to keep it reviewable independently.

[nastya236]

Thanks for flagging this! This one can be closed when we merge #1198.
Closing this PR for now.