Masked fill: slow operations compared to tch for small models

This is a follow up to: https://github.com/huggingface/candle/issues/2583

In our use case, the masking operations are quite a bit slower in candle vs tch: ~150 μs vs ~40 μs using a batch size of 200.

I’ve created a minrep repo demonstrating the same for various batch sizes, please see: https://github.com/JHucker/candle_mask_minrep

To make sure I was using candle masked-ops correctly, I followed the same mask creation and application from this [example](https://github.com/huggingface/candle/blob/e2b6b367fa852ed30ac532f8d77cd8479c7ed092/candle-transformers/src/models/mpt.rs#L278).

Thanks in advance, appreciate any help or feedback with regards to this.