RF scaled_dot_product_attention

[albertz]

Add scaled_dot_product_attention as a function to RF, and use it in our attention code. (Does this also work with RelPosSelfAttention?)

In case of PyTorch, wrap torch.nn.functional.scaled_dot_product_attention. That should be much more compute and memory efficient compared to the direct implementation. It uses FlashAttention or potentially a number of other efficient kernels. (Although on older GPUs, probably not. See my question on PyTorch discussion forum.)