Forcing deterministic behavior in Fabric

Nobody has claimed this yet.

### Description & Motivation

PyTorch has a utility to force deterministic algorithms via

```py
torch.use_deterministic_algorithms(True)
torch.backends.cudnn.benchmark = False
```

This can also be conveniently added in the Trainer:

```py
Trainer(deterministic=True)
```
(which also deactivates benchmark).

Fabric doesn't have this yet.

### Pitch

**Option 1: Init argument**

Introduce `Fabric(deterministic=bool|str)` with identical implementation/behavior as in Trainer.
Pro:
- Easy to discover

Con:
- Cannot disable it later on unless I create a new Fabric object. Maybe I want to disable it during training but enable it during validation/test/prediction.

**Option 2: Functional call**

Introduce a simple function `deterministic(bool|str)`.
Pro:
- Can be used anywhere you want, disable and enable on the fly

Con:
- Maybe not as easy to discover, but we can show it in examples

**Option 2: Both**
Simply do both options for flexibility.

### Alternatives

_No response_

### Additional context

_No response_

cc @borda @carmocca @justusschock @awaelchli

Start by reviewing the existing Trainer(deterministic=True) behavior and its handling of torch.use_deterministic_algorithms and cudnn.benchmark. Decide whether Fabric should expose an initialization argument, a functional call, or both; done means the selected interface provides equivalent deterministic behavior and supports the intended enable/disable use cases.

Written by the indexing model from the issue text.