Fixed attention precision not being cast in attention_pytorch and others functions. This led to the functions not being able to respect the `--dont-upcast-attention` flag.
change calls to `.float()` to `.to(dtype=torch.float32)` in several locations, as it profiles much faster.
removed unneeded check for `attn_precision == torch.float32`, as the change from `.float()` to `.to(dtype=cast_to_type)` does not cast or copy if `tensor.dtype == cast_to_type`
8299ebdaae