Adopt pytorch-directml as accelerator

Description & Motivation

i have tried to custome Accelerator as the code

class DML_Accelerator(Accelerator):
    def setup_device(self, device: torch.device) -> None:
        pass
    def setup(self, trainer: "pl.Trainer") -> None:
        pass
    def get_device_stats(self, device: _DEVICE) -> Dict[str, Any]:
        return {}
    def teardown(self) -> None:
        pass
    @staticmethod
    def parse_devices(devices: Any) -> Any:
        return devices
    @staticmethod
    def get_parallel_devices(devices: Union[int, str, List[int]]) -> List[torch.device]:
        t = torch_directml.device
        return [torch_directml.device()]
    @staticmethod
    def auto_device_count() -> int:
        return torch_directml.device_count()
    @staticmethod
    def is_available() -> bool:
        return torch_directml.is_available()
    @classmethod
    def register_accelerators(cls, accelerator_registry):
        accelerator_registry.register(
            "dml",
            cls,
            description=f"GPU Accelerator - optimized for large-scale machine learning.",
        )

class test_model(pl.LightningModule):
    def __init__(self, *args: Any, **kwargs: Any) -> None:
        super().__init__(*args, **kwargs)
        self._model = densenet121() 
    def training_step(self, batch,batch_index, **kwargs: Any) -> Any:
        y = self._model(batch)
        return None
    def configure_optimizers(self) -> Any:
        return None
    
class testSet(Dataset):
    def __init__(self) -> None:
        super().__init__()
        
    def __getitem__(self, index) -> Any:
        return torch.rand(3,224,224)
    
    def __len__(self):
        return 100

if __name__ == '__main__':
    acc = DML_Accelerator()
    trainer = pl.Trainer(accelerator=DML_Accelerator(),max_epochs=20)
    loader = (
        DataLoader(testSet(),batch_size=12,num_workers=8,drop_last=True),
    )
    trainer.fit(test_model(),*loader)

but the model style use the device 'cpu',not the directml after i pass the DML_Accelerator to the trainer.the pytorch_drectml module can work successfully.

Pitch

No response

Alternatives

No response

Additional context

No response

cc @borda @carmocca @justusschock @awaelchli

Start with the Accelerator entry point and Trainer's accelerator handling, then compare the supplied DML_Accelerator methods with existing accelerator implementations. Done means a supported pytorch-directml accelerator is selected by Trainer and the model runs on DirectML rather than CPU, with device availability and selection covered.