Resolve training type plugin when passed with Accelerator

[kaushikb11]

🐛 Bug

To Reproduce

trainer = Trainer(accelerator=GPUAccelerator(precision_plugin=PrecisionPlugin(), training_type_plugin=DDPPlugin()), gpus=4)

At the moment, when training type plugin is passed with Accelerators, attributes such as parallel_devices, cluster_environment and sync_batchnorm are not set to the training plugin and leads to errors.

Expected behavior

trainer = Trainer(accelerator=GPUAccelerator(precision_plugin=PrecisionPlugin(), training_type_plugin=DDPPlugin()), gpus=4)

# should be equivalent to

training_type_plugin.parallel_devices == [torch.device("cuda", i) for i in self.parallel_device_ids]
training_type_plugin.cluster_environment == LightningEnvironment()
training_type_plugin.sync_batchnorm == False

[awaelchli]

Thanks @kaushikb11
This fix would only go into 1.5.x and partially apply to master, because #10416 would change the requirements slightly here as things get inverted.

[kaushikb11]

Yup @awaelchli, I did check #10648 out before going ahead with the PR. Thanks!

[ananthsub]

@four4fish @kaushikb11 is this solved by the accelerator connector rewrite?

[four4fish]

@ananthsub yes it's solved by the rewrite and the follow ups
This is the last one #12105