Should Fabric have a `.mark_step()`?

Nobody has claimed this yet.

### Description & Motivation

In [#265](https://github.com/Lightning-AI/lit-llama/pull/265), the inference code gets sped up by inserting the XLA mark_step() calls. Should Fabric provide an accelerator/strategy-agnostic method for that?

### Pitch

It could be implemented directly like this in the Fabric class:

```py
def mark_step(self):
if self.device.type == "xla":
import torch_xla.core.xla_model as xm

xm.mark_step()
# for all other accelerators, this is a no-op
```

The convenience method enables you to call

`fabric.mark_step()` directly in the loop and no code changes are required once switching to another accelerator or strategy.

### Alternatives

The above proposes to add the logic directly to the Fabric object since it is very simple and concerns only the XLA strategy and to keep the base strategy interface lean. If this turns out to generalize and be required for other accelerators in the future, we could then decide to add the interface to the strategy base class.

### Additional context

_No response_

cc @borda @carmocca @justusschock @awaelchli

Start with the Fabric class and the accelerator or strategy handling described in the issue. Check the proposed torch_xla mark_step() behavior and determine how the method should behave for non-XLA accelerators; done means the Fabric-level call works for XLA and is a no-op elsewhere.

Written by the indexing model from the issue text.