fix(layers): Fix compute_dtype handling for complex-valued inputs in Dense layer

[harshaljanjani]

Fixed incompatibility between compute_dtype and complex-valued inputs in the Dense layer (issue #20278). The layer previously did not properly handle dtype promotion when mixing compute_dtype with complex inputs.

Key changes:

• Add proper dtype promotion in Dense.call() using backend.result_type()
• Cast inputs, kernel and bias weights to the promoted dtype.
• Add complex number initialization support via initializer_for_complex wrapper
• Add comprehensive test suite for complex number handling

This ensures the Dense layer correctly handles:

• Complex-valued inputs with different compute_dtype settings
• Mixed real and complex computations
• Complex weight initialization
• Complex number training

Tests added verify compatibility across different backends (TensorFlow, JAX, PyTorch) and complex number use cases.

[codecov-commenter]

Codecov Report

Attention: Patch coverage is 72.72727% with 15 lines in your changes missing coverage. Please review.

Project coverage is 82.01%. Comparing base (9c8da1f) to head (d8ca6cd).
Report is 3 commits behind head on master.

Files with missing lines	Patch %	Lines
keras/src/random/random.py	62.50%	7 Missing and 5 partials ⚠️
keras/api/_tf_keras/keras/random/__init__.py	0.00%	2 Missing ⚠️
keras/src/layers/core/dense.py	94.11%	0 Missing and 1 partial ⚠️

Additional details and impacted files

@@           Coverage Diff           @@
##           master   #20823   +/-   ##
=======================================
  Coverage   82.01%   82.01%           
=======================================
  Files         559      559           
  Lines       52291    52348   +57     
  Branches     8084     8096   +12     
=======================================
+ Hits        42884    42931   +47     
- Misses       7431     7435    +4     
- Partials     1976     1982    +6     

Flag	Coverage Δ
keras	81.82% <72.72%> (+<0.01%)	⬆️
keras-jax	64.29% <72.72%> (+0.02%)	⬆️
keras-numpy	59.03% <72.72%> (+0.03%)	⬆️
keras-openvino	29.83% <32.72%> (+<0.01%)	⬆️
keras-tensorflow	64.82% <72.72%> (+0.01%)	⬆️
keras-torch	64.28% <72.72%> (+0.09%)	⬆️

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]

Complex value support must be added for other layers as well. Please let me know if the changes made to Dense should be extrapolated to the other layers, as I intended this as a proof of concept. The issue will only be fully resolved when all layers support complex values.
Edit: I've tried to implement fixes solely dtypes.py without changing the layer implementation, but without fixing the initializer and the call() methods, it's tricky to implement it, since the previous PR (cited in the issue) seems to have already implemented the type promotion very well.

[harshaljanjani]

Just as an aside, I faced torch-xla CI issues in this PR as well, which were resolved by merging keras-master as of January 30, 2024.

[harshaljanjani]

Update: Support for complex values in layers is no longer under consideration, but this PR could serve as a way to subclass from layers.Layer and add complex value support if needed.