[PyTorch] Update to 2.1.2

[HGuillemet]

Included in this PR:

• Add missing mapping for std::array<CUDACachingAllocator::Stat>
• Update to PyTorch 2.1.2

[saudet]

Good to merge?

[HGuillemet]

I think so, but I wonder:
Shouldn't we add a preload property in torch_cuda so that people running with the cuda presets and not a system-wide cuda installed have the libraries extracted, without explicitly loading the cuda classes ?
Something like:

  preload = {"cudart@.12", "cublasLt@.12", "cublas@.12", "cudnn@.8", "nvToolsExt@.1", "cusolver@.11", "cufft@.11", "curand@.10", "cusparse@.12"}

[saudet]

That's already in presets/torch.java: https://github.com/bytedeco/javacpp-presets/blob/master/pytorch/src/main/java/org/bytedeco/pytorch/presets/torch.java#L147

[HGuillemet]

Ok. I forgot this.
So GtM.