Enhancement of one-shot NAS (v2.9)

[ultmaster]

Description

Features

• Add export_probs to monitor the architecture weights.
• Rewrite configure_optimizers, functions to step optimzers / schedulers, along with other hooks for simplicity, and to be compatible with latest lightning (v1.7).
• Support gradient clipping in DARTS.
• Support arbitary arguments in random one-shot, because it's automatic optimization.
• Support +1 (none) in differentiable cell, to mimic the input choice (similar to what DARTS did).
• Reimplement ProxylessNAS with forward/backward hooks. This increases its robustness and applicability.

Bug fixes

• Fix a misuse of lambda expression in NAS-Bench-201 (used in tests).
• Fix the gumbel temperature (the original one doesn't work at all).
• Fix an issue that formats tensors in weighted_sum, making it very slow.
• Fix the label sharing in differentiable strategies. The original one wasn't sharing at all.
• Fix the memo reuse in exporting differentiable cell.

Misc

• Add a more friendly error message when no architecture weights are available.
• Added some more explanations in one-shot documentation: manual optimization, some extra details in one-shot supermodules, typo fixes.

Test Options

• [ ] fast test
• [ ] full test - HPO
• [ ] full test - NAS
• [ ] full test - compression

Checklist

• [ ] test case
• [ ] doc

How to test

In another PR.