Reapply torch.compile in Fabric.setup()

[awaelchli]

What does this PR do?

Alternative implementation to #19192

---

Part of #17250

When you do

model = torch.compile(orig_model)
model = fabric.setup(model)

today, then you end up with

FabricModule(StrategyWrapper(OptimizedModule(orig_model)))

where StrategyWrapper represents the wrapper applied by the strategy, e.g., FSDP, DDP, DeepSpeedEngine, etc.

With the new feature enabled in this PR, you will now get:

FabricModule(OptimizedModule(StrategyWrapper(orig_model)))

Note how the optimized module gets applied outside the strategy wrapper, but inside the fabric module router.
It is important to have the fabric module outside the optimized module, because it contains only convenience routing and precision management that does not need to be compiled (nor can it be without graph breaks atm). At the same time, the FSDP/DDP specific operations can still be captured properly by torch compile, and how it is intended to be wrapped in PyTorch.

This feature is marked as experimental because we first need to test its application in the wild.

Follow-up work:

• Verify it in the wild
• Decide on the default
• Document it in the compile-guide
• Do the same in Lightning Trainer

---

📚 Documentation preview 📚: https://pytorch-lightning--19280.org.readthedocs.build/en/19280/

cc @Borda @carmocca @justusschock @awaelchli