[FSDP] Allow nested FSDP wrapper to use different mixed precision (#90523)
The main change is to move `args` and `kwargs` dtype convertion
from `_root_pre_forward` to `_pre_forward`, so that every
FSDP has a chance to apply its own precision.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/90523
Approved by: https://github.com/awgu, https://github.com/rohan-varma
