[NEEDS TESTING] Update dependencies

[FWao]

This PR updates Python to 3.13.

To avoid building causal-conv1d & mamba-ssm we hard-coded wheels for linux x86_64 and aarch64 - specific for torch 2.10, CUDA 13.0 and Python 3.13.

Unfortunately as no current flash-attn wheels exist, we must build flash-attn ourselves which can take a long time. To speed up the installation, the gpu dependency group is now installed without flash_attn! Users must use gpu_all to build flash_attn to use some of the models!

Some smaller changes were necessary to remain compatibility with recent timm & torch version!

This PR was tested on a DGX Spark!

[mducducd]

I have run tests on real data, and so far no major regressions. The core workflows completed as expected.
With gpu_all, flash-attn installation takes quite a long time, but that seems to be the practical trade-off here.

[FWao]

I have run tests on real data, and so far no major regressions. The core workflows completed as expected. With gpu_all, flash-attn installation takes quite a long time, but that seems to be the practical trade-off here.

Thank you!

I think we could provide a repo where we provide pre-built flash-attn wheels for Python 3.13 / torch 2.10 / cuda 13 for x86_64 and aarch64. Maybe similar to here.