Dtensor policy v1 and v2 future plan

[joyang-nv]

There are several issues about dtensor policy v1 and v2 now.

1. There are problems for v2 (NeMo Automodel to support some diffusion models.)
2. Researchers enjoy v1 merits: fully HF native and transparency.

But there are other issues from SW perspective:

1. NeMo Automodel will be much more powerful and feature rich built on top of FSDP.
2. There are many duplicated codes in v1 and v2 which will make maintenance more difficult.

With many discussions between researching we had some conclusions:

1. Keep v1 till they feel comfortable to deprecate.
2. Increase Automodel support for gapped features.
3. Automodel will have HF native fallback mode and this will be supported with in v2 dtensor policy worker.
4. Our refactor works on v2 policy worker will also apply to v1 to make code cleaner on both side.
5. There will be divergent features between v1 and v2 in future releases including EP, PP, HSDP, CP + Seq pack via TE.