Improve performance docs

[lvwerra]

What does this PR do?

As discussed with @stas00 this PR does the following things:

• adds files for missing sections so contributors can better find the place to add content.
• adds a disclaimer at the beginning stating that a lot of general training information is in the single GPU training section
• fixes the link of CPU inference

Looking at the ToC I was wondering whether we should add subsections like it is done for the tasks to make the main ToC a bit slimmer. Otherwise we have the main performance docs (the entry point) there plus all the other sections (~8-10). What do you think?

Before submitting

• This PR fixes a typo or improves the docs (you can dismiss the other checks if that's the case).
• Did you read the contributor guideline,
  Pull Request section?
• Was this discussed/approved via a Github issue or the forum? Please add a link
  to it if that's the case.
• Did you make sure to update the documentation with your changes? Here are the
  documentation guidelines, and
  here are tips on formatting docstrings.
• Did you write any new necessary tests?

[HuggingFaceDocBuilderDev]

The documentation is not available anymore as the PR was closed or merged.

[sgugger]

Thanks for working on this. Could we add at least one line in each of those files telling the reader "Coming soon"?
Definitely yes for grouping all of those guides in one subsection, like the Tasks one.

[lvwerra]

Sounds good @sgugger, I added the "Coming soon" and restructured the ToC.

[stas00]

Communications are difficult.

what I proposed is to add normal documents - not "Coming Soon"

inside those documents they should all explicitly point the user to read the first document that is already filled out and that we will expand this other currently mostly empty doc with the missing material.

In other words e.g,, perf_infer_gpu_many.mdx should have:

1. read perf_train_gpu_one.mdx first
2. this document will be completed soon.

If the docs remain "Coming Soon" nobody will read them and will miss out on the already rich performance docs we have.

And the impetus for this change was that we went from a complete solution - all performance notes in one doc, to very incomplete solution, making it look like we only have advise for those with 1 gpu and doing training.

The original discussion back in winter was that all performance docs will be filled out, but it was dropped after the first document and no signs of new docs coming any time soon. So this proposal was my attempt to rescue the situation.

[flozi00]

Typo ? :)

[lvwerra]

Indeed, thanks!

[lvwerra]

Ok I added references to each document and a small text explaining what will come there. Is that what you had in mind @stas00?

[stas00]

With a few suggestions to fix this looks good.

In particular I added pointers to 1-gpu train from infer docs since most of it applies and the current infer files are almost empty and don't have the essentials.