Fix SetTransformer kwargs handling

[arrjon]

This pull request introduces improvements to keyword argument handling in SetTransformer (please check if this fine) and adds tests for the ContinuousApproximator sampling functionality. Before calling workflow.sample with additional kwargs, e.g. method='euler' for FlowMatching, was failing with the SetTransformer.

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.

Files with missing lines	Coverage Δ
bayesflow/networks/transformers/mab.py	96.77% <100.00%> (ø)
bayesflow/networks/transformers/set_transformer.py	93.75% <100.00%> (ø)

... and 3 files with indirect coverage changes

[stefanradev93]

I was surprised that the existing filter in ContinuousApproximator (L525-527) is not sufficient:

summary_outputs = self.summary_network( summary_variables, **filter_kwargs(kwargs, self.summary_network.call) )

Perhaps removing the unnecessary **kwargs argument from all attention helpers (e.g., MultiHeadAttentionBlock) solves it cleanly?

Perhaps it is because the call method itself accepts kwargs, so the argument gets propagated down to the multihead attention block.

[arrjon]

Perhaps removing the unnecessary **kwargs argument from all attention helpers (e.g., MultiHeadAttentionBlock) solves it cleanly?

I can have a look.

Perhaps it is because the call method itself accepts kwargs, so the argument gets propagated down to the multihead attention block.

That is indeed the problem.

[arrjon]

self.attention_blocks = keras.Sequential()
this part did not accept kwargs at all, so I removed them entirely