Upgrade transformers to 5.x and other dependencies

[dxqb]

There are some major upgrades of dependencies:

• transformers from 4.x to 5.0
• huggingface_hub from 0.34.x to 1.3.x
• Torch 2.10
• python 3.13
• CUDA 13.0

This PR upgrades all dependencies to their latest version (except ROCM).
huggingface_hub was held back by transformers so far. With transformers 5.0, huggingface_hub gets a major upgrade.

This is why I've enabled XET in this PR. Previous problems with XET might not apply anymore - to be tested.
If XET works well, #770 could be reverted because you don't need to download multiple files anymore to get full speed.
Would close #949

[dxqb]

updating to CUDA 13 might be too early:

/OneTrainer/venv/lib/python3.12/site-packages/torch/cuda/__init__.py:184: UserWarning: CUDA initialization: The NVIDIA driver on your system is too old (found version 12090). Please update your GPU driver by downloading and installing a new version from the URL: http://www.nvidia.com/Download/index.aspx Alternatively, go to: https://pytorch.org to install a PyTorch version that has been compiled with your version of the CUDA driver. (Triggered internally at /pytorch/c10/cuda/CUDAFunctions.cpp:119.)

[dxqb]

CUDA 12.8 instead. Even many RTX 5090 clouds are still on drivers not supported by CUDA 13

[O-J1]

Self note: Tested and the scalene upgrade breaks on windows. Need to look into it

[dxqb]

Self note: Tested and the scalene upgrade breaks on windows. Need to look into it

or downgrade to that version, but leave a comment that scalene must not be upgraded. which version?

[O-J1]

Unfortunately have to revert.

Findings:

1. This PR is bricked in all versions of python we can support thanks to Transformers v5. Migration Guide here
2. The chosen numpy and scipy variants drop support for python 3.10
3. We'd have to downgrade transformers and huggingface-hub an 'old-new' version (still published in 2026 but maitenence versions instead of breaking)
4. Scalene also had breaking changes, pin at 1.5.55 or upgrade commands.

[dxqb]

torch 2.10 alone first #1266 because this PR requires some more work

[hameerabbasi]

On ROCm: It's only the first launch for any kernel that's slow on 2.10. It's actually a plus since most things are compiled on the fly, meaning the binary is actually smaller. So let's keep ROCm.

[dxqb]

when we do, #1346 can be reverted

[hameerabbasi]

4. Scalene also had breaking changes, pin at 1.5.55 or upgrade commands.

We're using only the parts of Scalene that didn't break change their API with 2.x, but I haven't tested it actually works.