Add MUSA graph-safe op overrides and torch.load map_location translation

[froststeam]

Summary

• translate CUDA device strings passed to torch.load(..., map_location=...) to MUSA devices on MUSA platforms
• add graph-safe MUSA PrivateUse1 overrides for aten::multinomial, aten::log, and aten::log_
• bump package references from 0.1.83 to 0.1.84

Notes

• map_location dict source keys are kept unchanged so serialized CUDA storage tags still match; only target devices are translated.
• The new op overrides are gated by the existing TORCHADA_DISABLE_OP_OVERRIDE_<OP_NAME>=1 mechanism.
• MUSA op tests use torch.musa.is_available() because torch.cuda.is_available() remains false in this torchada CUDA-compatible setup, while CUDA device strings still map to MUSA tensors.

MUSA Test

Environment:

• host: 10.20.35.93, GPU: 2 (MUSA_VISIBLE_DEVICES=2, MTHREADS_VISIBLE_DEVICES=2)
• image: registry.mthreads.com/mcconline/inference/sglang:v0.5.12.post1-ph1-4.3.5-torch2.9.0-latest
• torch: 2.9.0
• torch.version.musa: 40305
• torch_musa: 2.9.0+571893f
• torchada from PR branch: 0.1.84
• torchada.is_musa_platform() == True
• torch.musa.is_available() == True, torch.musa.device_count() == 1
• torch.empty(1, device="cuda").device == musa:0

Command:

cd /tmp/torchada-pr105
MUSA_VISIBLE_DEVICES=2 MTHREADS_VISIBLE_DEVICES=2 PYTHONPATH=src \
  pytest -q tests/test_device_strings.py tests/test_log.py tests/test_multinomial.py

Result:

51 passed in 9.88s

Local Check

• git diff --check