Fix numeric instability in LayerNormalization

[james77777778]

When implementing SD3, I found that LayerNormalization is numeric unstable when dtype is set to float16.
This PR upcasts float16/bfloat16 to float32 for the computation.

I used backend.result_type(..., "float32") to handle higher precision dtypes, such as float64.

[codecov-commenter]

Codecov Report

All modified and coverable lines are covered by tests ✅

Project coverage is 79.35%. Comparing base (e7b5a5d) to head (39f29b7).

Additional details and impacted files

@@            Coverage Diff             @@
##           master   #20257      +/-   ##
==========================================
- Coverage   79.35%   79.35%   -0.01%     
==========================================
  Files         501      501              
  Lines       47642    47637       -5     
  Branches     8768     8765       -3     
==========================================
- Hits        37808    37803       -5     
  Misses       8054     8054              
  Partials     1780     1780              

Flag	Coverage Δ
keras	79.20% <100.00%> (-0.01%)	⬇️
keras-jax	62.49% <100.00%> (-0.01%)	⬇️
keras-numpy	57.61% <100.00%> (+<0.01%)	⬆️
keras-tensorflow	63.88% <100.00%> (-0.01%)	⬇️
keras-torch	62.54% <100.00%> (+<0.01%)	⬆️

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.