chore(docker): update base image to new cuda & torch

[ndeybach]

Updates the base image to newer cuda and pytroch to better support newer gpu architectures.

• from "pytorch:2.4.0-cuda12.1-cudnn9-runtime" to "pytorch/pytorch:2.9.0-cuda12.8-cudnn9-runtime"

[ndeybach]

@Vincentqyw , thanks for merging my other PR just now.

For this one I think that from my november memory I later encountered a crash / incompatibility when testing one model (that was build as research code and never upgraded beyond 12.1 ... so there may be incompatibilities.

I don't know how you would like to handle updates to newer cuda versions (which new models do need though for some aspects and acceleration / optimizations when benchmarking them on new arch)...

[Vincentqyw]

Hi @ndeybach, to be honest, I don't have a perfect solution for version management either; it's mostly been trial and error on my end.
Regarding new algorithms, my usual approach is to check their requirements.txt and compare it with imcui's current dependencies. If there are no new mandatory additions, I prefer to keep the status quo. As of now, I haven't encountered a pressing need to upgrade PyTorch or CUDA.
My philosophy is that since most open-source algorithms are built upon earlier works (which have already gone through extensive adaptation), the amount of compatibility work left for me is actually quite manageable.