[deepspeed doc] install issues + 1-gpu deployment

[stas00]

This PR extends the DeepSpeed/FairScale integration documentation to:

• add extensive general troubleshooting for CUDA-extensions (applies to fairscale, deepspeed, apex or any other python pytorch extension with CUDA C++ code) - these are very likely to be encountered by our users - all notes are based on my first hand encounters with these issues - 2 of which I run into yesterday while trying to build fairscale and deepspeed on Sylvain's hardware which he let me use to run the recent benchmarks. so I figured others are likely to have similar issues and neither fairscale nor deepspeed have these documented anywhere.
• adds deployment for 1 gpu DeepSpeed notes
• reformats sub-headers so that it's easier to link to specific sections

@sgugger

[LysandreJik]

Great! Thanks for writing down the fixes to the issues you fixed, that's helpful!

[sgugger]

Thanks a lot for expanding the doc! Just left some nits (mostly about the correct case of some names).

[stas00]

Thank you for your awesome suggestions and tweaks - all done.