[RFC] Deprecate `auto_select_gpus` from Trainer constructor

[daniellepintz]

Proposed refactoring or deprecation

Deprecate auto_select_gpus from Trainer constructor:
https://github.com/PyTorchLightning/pytorch-lightning/blob/7a9a08c5d3ca4699fb439f691c40e1320b37507a/pytorch_lightning/trainer/trainer.py#L209-L212

Motivation

Users can achieve the same functionality by setting the accelerator and devices flags - we can simplify by only having one way of doing this.

cc @justusschock @awaelchli @rohitgr7 @tchaton @kaushikb11 @Borda

[kaushikb11]

@daniellepintz We do have users using it. For instance, #10456.
It could be pretty convenient when users are running multiple experiments on bunch of GPUs.

[daniellepintz]

@kaushikb11 can this same functionality not be achieved with accelerator and devices flags?

[rharish101]

@daniellepintz Currently, I think that if we want 3 available GPUs, then we need to specify devices=3 and accelerator='gpu'. However, this means that GPUs [0, 1, 2] are selected, even if one of them is fully used by another process:
https://github.com/PyTorchLightning/pytorch-lightning/blob/ffb1a754af5301608c71955c12468216e920af51/pytorch_lightning/utilities/device_parser.py#L76

However, with auto_select_gpus, it chooses 3 GPUs that are available. This will be [0, 1, 2] if all three are available, else it will also check GPUs 3, 4, etc. until it finds 3 available GPUs.

[ananthsub]

n00b question: why isn't this always enabled, such that if someone specifies Trainer(accelerator="gpu", devices=3) Lightning automatically chooses 3 non-reserved GPUs?