Refactoring: turn the attribute _return_attention_scores into an argument

[apehex]

I have a small refactoring for the class MultiHeadAttention 👾

I propose to replace the private attribute _return_attention_scores with an extra argument in _compute_attention.

This would make subclassing more straightforward and avoid cases where the attention scores are None.
This issue shows in CachedMultiHeadAttention on KerasHub.

The rationale is explained in more details in #20802

[codecov-commenter]

Codecov Report

All modified and coverable lines are covered by tests ✅

Project coverage is 82.01%. Comparing base (90568da) to head (4c5557e).

Additional details and impacted files

@@            Coverage Diff             @@
##           master   #20803      +/-   ##
==========================================
- Coverage   82.01%   82.01%   -0.01%     
==========================================
  Files         557      557              
  Lines       52016    52014       -2     
  Branches     8037     8037              
==========================================
- Hits        42659    42657       -2     
  Misses       7403     7403              
  Partials     1954     1954              

Flag	Coverage Δ
keras	81.83% <100.00%> (-0.01%)	⬇️
keras-jax	64.22% <100.00%> (-0.01%)	⬇️
keras-numpy	58.96% <0.00%> (-0.01%)	⬇️
keras-openvino	29.89% <0.00%> (+<0.01%)	⬆️
keras-tensorflow	64.78% <0.00%> (-0.01%)	⬇️
keras-torch	64.17% <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.

[fchollet]

Right, this is a good change. Thanks!