[torchlib] Fix sdpa dtype in attn_mask

[titaiwangms]

Discovered in benchmark that the op.Where generates fp32 output when the whole model is set to fp16.

[codecov]

Codecov Report

Attention: Patch coverage is 0% with 3 lines in your changes missing coverage. Please review.

Project coverage is 69.55%. Comparing base (a998e5d) to head (64a23d8).
Report is 2 commits behind head on main.

Files with missing lines	Patch %	Lines
onnxscript/function_libs/torch_lib/ops/nn.py	0.00%	3 Missing ⚠️

Additional details and impacted files

@@            Coverage Diff             @@
##             main    #2445      +/-   ##
==========================================
- Coverage   69.55%   69.55%   -0.01%     
==========================================
  Files         210      210              
  Lines       26130    26132       +2     
  Branches     2722     2722              
==========================================
  Hits        18176    18176              
- Misses       7018     7020       +2     
  Partials      936      936              

☔ View full report in Codecov by Sentry.
📢 Have feedback on the report? Share it here.