Add support for einsum operation to pytorch parser (requires 1116)

[JanFSchulte]

Builds on Chang's keras v3 PR #1116 and exposes the new einsum implementation through the pytorch parser. pytorch doesn't have an an equivalent to EinsumDense but allowing the use of einsum operations in some custom model would still be useful.

Type of change

• New feature (non-breaking change which adds functionality)

Tests

Added 2 use cases (outer product and batch matrix multiplication to the pytests, works without issues.

Checklist

• I have read the guidelines for contributing.
• I have commented my code, particularly in hard-to-understand areas.
• I have made corresponding changes to the documentation.
• My changes generate no new warnings.
• I have installed and run pre-commit on the files I edited or added.
• I have added tests that prove my fix is effective or that my feature works.

[JanFSchulte]

Now that #1116 is merged, this can also be added.

[calad0i]

It is preferred to use hls4ml.utils.einsum_utils._validate_einsum_expr here, something like:

inp_shape0 = input_shapes_tmp[0][1:]
inp_shape1 = input_shapes_tmp[1][1:]

layer['equation'], layer['out_shape'] = _validate_einsum_expr(node.args[0], inp_shape0, inp_shape1

In this way wildcard dimensions (...) will be fixed to indices and explicit torch call will not be required.
Otherwise all looks good.

[JanFSchulte]

Thanks for the suggestion, implemented.

[calad0i]

Ok, actually I didn't expect the batch dimension when implementing this, and I explicitly striped off the batch dim in the keras impl. However, indeed setting batch dimension = 1 is functionally equivalent and should always work.

Merging after test pass.