Refactor/torch autocast encapsulate global state

[nathon-lee]

refactor: replace bare global vars in torch_autocast with _AutocastState

TORCH_AUTOCAST_INITIALIZED and TORCH_AUTOCAST_DTYPE were module-level
globals mutated via global statements inside init_autocast_params().
This pattern is fragile: it is invisible to type checkers, prevents
isolation between multiple engine instances, and makes the state harder
to reset in tests.

Replace them with a private _AutocastState dataclass instance
_autocast_state. The public API (is_autocast_initialized,
get_autocast_dtype) is unchanged, so no call sites are affected.

---

fix: store autocast state per-engine to support multiple engine configs

Previously, _autocast_state was a module-level singleton in
torch_autocast.py. When a second DeepSpeed engine called
init_autocast_params(), it would overwrite the first engine's dtype
and initialized state, making it impossible to run two engines with
different autocast configurations concurrently.

Fix by attaching _AutocastState directly to the engine instance
(engine._autocast_state). Update is_autocast_initialized() and
get_autocast_dtype() to accept an engine argument. For ZeRO
optimizers (which hold no engine reference), switch from the global
state query to the per-parameter has_comm_dtype() check; parameters
are already stamped by their own engine inside init_autocast_params(),
so isolation is automatic.

[tohtana]

Hi @nathon-lee, thank you for opening this PR!
_autocast_state is still global and doesn't seem support different configs for multiple engines. Did I misunderstand something?

[nathon-lee]

**tohtana **

@tohtana thank you, good catch, — I still need to make one more change.