[FSDP2, Do not merge] Refactor

[S1ro1]

A somewhat big PR, started as enabling fp8 with torchao, ended up realising fsdp2 is too b*tchy about order of operations, so went for a bigger refactor and few changes, going over them biggest -> smallest

1. Move whole FSDP2 preparation to accelerator._prepare_fsdp2 such as in other cases as deepspeed - this enables bigger freedom in the future, when composing FSDP2 with other features. This has proved to be beneficial already where this enabled torchao fp8 support

2. FSDP2 specific compile - we now compile after model converters (AC/FP8) but before FSDP2, this fixes compile issues and is actually how torchtitan does it

3. Move optimiser parameter switch to be FSDP2 specific - this can now reside in the method mentioned above, enabling us to do simpler fsdp2 specific things. Now we canonicalise names for old and new params (not 100% sure about the reach of this, but proved to be ok with compile + AC + fp8)

4. Closely related to 2, we now do a bit different order of operations, i.e. applying activation checkpointing before compile, that before fully_shard

5. By accident, this PR also fixes FP8 with torchao, so the changes are included in there (should be minor)

I have been extensively testing this for a ~week, so I'm quite confident, though you never know. If @winglian had any time to test if this doesn't break anything downstream, would very much appreciate it.

[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.