[Offloading] [Bugfix] Fix disk offloading of models with explicit tensor dtypes

[kylesayrs]

Purpose

• Fix bug where DSV4 would load tid2eid with the incorrect precision, leading to incorrect hash gate values
  • Casting from int64 to bfloat16 was causing tid2eid to round to out-of-bounds values, not to mention that the values were incorrect

Changes

• Instead of using model.dtype to determine the load dtype for disk-offloaded weights, instead use the dtype of the associated meta tensor in the meta model. The meta tensor dtype is already the correct dtype, influenced by the local_torch_dtype load context
  • This dtype field is read by OffloadedWeightsLoader here

Testing

• Added test_disk_onload_dtype

RUN_SLOW=1 python3 -m pytest tests/utils/test_modeling_utils.py::ModelUtilsTest::test_disk_onload_dtype -sx

Suggested Reviewers

@SunMarc @zucchini-nlp

[github-actions]

CI Dashboard: View test results in Grafana