fix(layers): Fix incorrect masked mean/variance in BatchNormalization layer

[harshaljanjani]

• Update masked moments calculation to properly account for broadcast dimensions when summing mask weights.
• Added test to verify broadcast mask handling produces zero-centered outputs.
• Fixes BatchNormalization gives incorrect output with masked inputs > 3 dimensions #19848.

[codecov-commenter]

Codecov Report

All modified and coverable lines are covered by tests ✅

Project coverage is 82.01%. Comparing base (734cd03) to head (1db8cc0).
Report is 1 commits behind head on master.

Additional details and impacted files

@@            Coverage Diff             @@
##           master   #20815      +/-   ##
==========================================
- Coverage   82.01%   82.01%   -0.01%     
==========================================
  Files         558      559       +1     
  Lines       52182    52291     +109     
  Branches     8067     8084      +17     
==========================================
+ Hits        42795    42884      +89     
- Misses       7421     7431      +10     
- Partials     1966     1976      +10     

Flag	Coverage Δ
keras	81.82% <100.00%> (-0.01%)	⬇️
keras-jax	64.27% <100.00%> (+0.02%)	⬆️
keras-numpy	59.00% <100.00%> (+0.02%)	⬆️
keras-openvino	29.83% <100.00%> (-0.02%)	⬇️
keras-tensorflow	64.80% <100.00%> (+0.01%)	⬆️
keras-torch	64.19% <100.00%> (+0.03%)	⬆️

Flags with carried forward coverage won't be shown. Click here to find out more.

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

[harshaljanjani]

The error is in the line in ops.mean() for the OpenVINO backend:
axis_const = ov_opset.constant(axis, dtype=Type.i32).output(0)
Until further information is available, I will mark the test case as skipped for the OpenVINO backend.
Edit: Fixed.

[fchollet]

Thanks for the fix!

[harshaljanjani]

@fchollet I know it's been a long PR with the RandomChoice and RandomApply layers, but it's my first month contributing, and I've learned a lot about src/layers/preprocessing during the migration process. It's alright if you don't want those layers in the codebase, but I'll fix the reviews and re-request a review anyway, thanks!
Just let me know if there's any project you're looking for contributors on, I’d love to learn and contribute!

[fchollet]

LGTM, thank you