What does this PR do?

Currently the torch.onnx.export tracer fails to export the ecapa-tdnn embedding model, giving the error described in #149760. This error occurs because an exception is part of the control flow of the model, used to check whether or not a given layer supports the length argument. When this is thrown, the tracer has already pushed the layer onto its stack so that, while the layer is not used, it sticks around until the end of the trace, where it is encountered instead of the root node, hence the key mismatch.

The following very minimal code triggers the failure. I have also successfully exported the embedding model to onnx with the dynamo exporter, including dynamic batch and audio sample dimensions.

import torch
from speechbrain.inference.speaker import EncoderClassifier

enc = EncoderClassifier.from_hparams(source='speechbrain/spkrec-ecapa-voxceleb').to('cpu')

torch.export.export(enc.mods.embedding_model, (torch.zeros(1, 1000, 80), torch.ones(1)))

Fixes #149760

PR review