Refactoring: turn the attribute _return_attention_scores into a method argument

[apehex]

Hey there! 👾

Issue

The class MultiHeadAttention is using a private attribute to decide whether or not to return the attention scores.

This behavior is not apparent from the signature of _compute_attention, because there are no corresponding argument.

So, if a sub-class fails to set _return_attention_scores in its custom call method, then _compute_attention never returns the scores.

For example, the child class CachedMultiHeadAttention from KerasHub uses _compute_attention and receives None as attention scores in all cases.

Additional context

The attribute _return_attention_scores is used in a single method, _compute_attention.

Proposition

IMHO the attribute _return_attention_scores should be removed.

Instead, _compute_attention should have an extra parameter return_attention_scores.
Then the call methods could just forward this value to _compute_attention explicitely.

I'll write a PR shortly if you find this relevant.