Fix torch.compile for timestep_embedding

[leeclemnet]

What does this PR do?

The use of np.log in timestep_embedding functions causes SAM3_3D inference to fail when torch_compile=True.

The reason is that when torch.compile traces this, dynamo sees -np.log(max_period) and tries to negate it through torch_function, which calls .neg() on the numpy float64 result — and numpy scalars don't have .neg(). Switching to math.log avoids the issue by returning a plain Python float.

Associated inference PR: roboflow/inference#2041

Type of Change

• Bug fix (non-breaking change that fixes an issue)

Testing

• I have tested this change locally
• I have added/updated tests for this change <-- no tests in tdfy currently

Test details:

Compilation with torch_compile=True succeeds after ~20min with torch_compile=True, model behaviour is not affected

Checklist

• My code follows the style guidelines of this project
• I have performed a self-review of my own code
• I have commented my code where necessary, particularly in hard-to-understand areas
• My changes generate no new warnings or errors
• I have updated the documentation accordingly (if applicable)

Additional Context