Convert BetterTranformer model - no support for _nested_tensor_from_mask

Hi,

I optimized a huggingface model with the BetterTransformer framework using [huggingface optimum.](https://huggingface.co/docs/optimum/bettertransformer/overview)
Here the simple code snippet:
```
from transformers import AutoModelForSequenceClassification
from optimum.bettertransformer import BetterTransformer
newModel = BetterTransformer.transform(baseline_model, keep_original_model=True)
```
BetterTransformer avoids unnecessary processing of padding tokens with nested tensors.

When I try to convert the optimized model I get the following error:

`_PyTorch convert function for op '_nested_tensor_from_mask' not implemented._`

Is there any way to fix the error? I haven't found any approach to implement the '_nested_tensor_from_mask' operation accordingly yet.

Thanks in advance!

Reproduce the failure using the BetterTransformer conversion snippet and the model-conversion path described in the issue. Start by locating how coremltools handles the reported `_nested_tensor_from_mask` operation; done means the optimized model converts without the reported unsupported-operation error and the behavior is covered by an appropriate conversion check.

Written by the indexing model from the issue text.