Add CLI to install dependencies

[tengomucho]

What does this PR do?

Working with TPU requires installing some dependencies that are as simple to install as adding a pip dependency in the main package. This is even more evident for Jetstream Pytorch, that needs to be installed from a git repository.
Previously I had run the actions manually, added a target on Makefile and a script for it.
@mfuntowicz suggested that might not be practical for someone having the optimum-tpu package but not the sources:
#99 (comment)

As a result, I added a script and integrated it as the optimum-tpu CLI to provide a simple command to install Jetstream Pytorch:

optimum-tpu  install-jetstream-pytorch 

Note that I also added an option to install pytorch xla. This is because in the future I might remove this dependency from optimum-tpu, since it is not required for inference when using Jetstream.

Before submitting

• Did you make sure to update the documentation with your changes?

[HuggingFaceDocBuilderDev]

The docs for this PR live here. All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update.

[dacorvo]

LGTM, thanks !

[mfuntowicz]

LGTM thanks for looking into see, can really make a big difference for the end user.

2 small nits otherwise 👌

[mfuntowicz]

Is 24.04 not supported yet?

[tengomucho]

No, we had the same problem last week on optimum, see here. Ubuntu latest comes with python 12, not supported by torch xla yet.

[mfuntowicz]

Should you source the version from torch directly? Seems very error prone / easy forgettable in the long run isn't it?

[tengomucho]

I was actually thinking in the long run I will try to remove the dependency altogether, so users could install optimum-tpu and then install torch cpu using this CLI: this will allow to avoid installing the cuda dependencies of the default torch package.