[dai123x]

What does this PR do?

Adds typing.overload signatures for Fabric.setup, Fabric.setup_optimizers and Fabric.setup_dataloaders, so the declared return type follows the number of arguments.

Fixes #21988

Fabric.setup previously returned Any, and the other two methods returned a union of the single- and the multi-object case, so type checkers lost the shape of the most common calls and every access on the result needed a cast or an explicit annotation (examples in #21988). Annotations only — no runtime behavior change.

call	before	after
fabric.setup(model)	Any	_FabricModule
fabric.setup(model, optimizer)	Any	tuple[_FabricModule, _FabricOptimizer]
fabric.setup(model, optimizer, scheduler=lr)	Any	tuple[_FabricModule, _FabricOptimizer, LRScheduler]
fabric.setup(model, opt1, opt2)	Any	tuple[_FabricModule | _FabricOptimizer | LRScheduler, ...]
fabric.setup_optimizers(opt)	union	_FabricOptimizer
fabric.setup_optimizers(opt1, opt2)	union	tuple[_FabricOptimizer, ...]
fabric.setup_dataloaders(loader)	union	DataLoader
fabric.setup_dataloaders(loader1, loader2)	union	list[DataLoader]

Design notes (kept as close to the sketch in #21988 as possible):

• The "multiple objects" overloads require at least two positional optimizers/dataloaders instead of a plain *args. Calls the runtime rejects (e.g. setup_optimizers() with no arguments) then fall through to the implementation signature instead of being typed with a shape the runtime never returns.
• setup(module=model) keeps working: module stays keyword-callable, while the optimizer arguments are positional-only, matching the runtime signature.
• The multi-optimizer union also covers the scheduler, which the runtime appends last when one is provided.

For reviewers

• Checked with mypy 1.20.0 (the version pinned in requirements/typing.txt): no new errors. A consumer snippet resolves every shape in the table above, verified with reveal_type.
• pytest tests/tests_fabric/test_fabric.py tests/tests_fabric/test_wrappers.py passes locally (Windows, CPU torch 2.12.0): 112 passed, 66 skipped (distributed/TPU skips), 0 failures.
• No runtime tests added since behavior is unchanged; no docs or CHANGELOG update needed for an annotation-only change.

Before submitting

• Was this discussed/agreed via a GitHub issue? Yes — Add overloads to Fabric.setup, setup_optimizers, and setup_dataloaders for precise return types #21988 describes the problem and sketches these overloads.
• Did you read the contributor guideline, Pull Request section?
• Did you make sure your PR does only one thing, instead of bundling different changes together?
• Did you make sure to update the documentation with your changes? (if necessary) — not necessary, annotations only
• Did you write any new necessary tests? (not for typos and docs) — no new runtime tests; behavior unchanged, typing verified with mypy as described above
• Did you verify new and existing tests pass locally with your changes?
• Did you list all the breaking changes introduced by this pull request? — none

No breaking changes — annotations only.