Change cached seq_len to int to enable compilation

[f0k]

First of all, thanks for the nicely reusable package!

With recent versions of PyTorch, torch.compile() fails with the error "RuntimeError: aten::copy() Expected a value of type 'Tensor' for argument 'src' but instead found type 'int'." on the line that does self.cached_freqs_seq_len.copy_(seq_len), where seq_len is an int.
Also it warns about a "Graph break from Tensor.item()" on the line that has (offset + seq_len) <= self.cached_freqs_seq_len.item().

This PR fixes both by changing cached_freqs_seq_len and cached_scales_seq_len from singleton int tensors to plain Python integers. Forgive me if I overlooked anything, but it seems to me that there is no benefit of having these values on the GPU?

[lucidrains]

@f0k thank you for the PR Jan! yes indeed, not sure why it was stored that way