perf(moe): avoid intra-op fanout while filling NVFP4 host banks

[earlvanze]

Change

Port the NVFP4 host-bank copy optimization to the current expert-piece loader. During CPU host-bank placement, temporarily use one PyTorch intra-op thread and restore the prior setting on success or failure. The scope is NVFP4 host banks; resident GPU packing and other expert formats retain their current threading.

The previous optimization lived in a model-specific loader, which upstream has since replaced with build_expert_banks. This patch applies at that shared placement point. It also isolates the loader optimization from the broader GLM-5.3 Flash PR (#292).

Evidence

On this host, 100 warmed 128 KiB CPU copy_ calls took 3.03–4.38 seconds with 12 intra-op threads and 0.0003 seconds with one thread. This is a microbenchmark of the copy operation, not a model startup or inference benchmark.

CUDA_VISIBLE_DEVICES='' PYTHONPATH=python /home/umbrel/.venv/bin/python -m pytest tests/moe/test_expert_bank_threading.py tests/moe/test_offload.py tests/moe/test_nvfp4_backends.py -q -m 'not needs_weights': 23 passed, 14 skipped (GPU-dependent on this host).

Focused Ruff F checks and git diff --check pass. The new test verifies bank contents, one-thread placement, and restoration after a packing error.