Cleaner API for utilizing all GPUs if available

[ChaiBapchya]

Description

Can we have a cleaner way of utilizing GPUs?

Current scenario

According to this [1]

from mxnet import npx
npx.set_np()
num_gpus = npx.num_gpus()
ctx = [mx.gpu(i) for i in range(num_gpus)] if num_gpus > 0 else [mx.cpu()]

Equivalent in PyTorch would be

device = torch.device("cuda" if torch.cuda.is_available() else "cpu")

Proposal

To make this easier for first time user of MXNet or (PyTorch user migrating to MXNet)
Can we have 2 changes

1. Additional API
   mx.gpu().is_available()
   or something on similar lines

Benefit - Adding this API would save user 3 lines of code and using npx.num_gpus() >1 is a round-about way of saying cuda.is_available()

Created a discussion on discuss forum for the same [2]

2. Additional boolean parameter
   mx.gpu(all_gpus=True)

Benefit - Adding all_gpus parameter is convenient than [mx.gpu(i) for i in range(num_gpus)]

References

1. https://gluon-cv.mxnet.io/build/examples_classification/transfer_learning_minc.html
2. https://discuss.mxnet.io/t/cuda-is-available-in-mxnet/5096

[nickguletskii]

I think it would be better to introduce a separate function called mxnet.all_gpus(): List[mxnet.Context], instead of adding a parameter to mxnet.gpu. This way, the return type of mxnet.gpu will remain mxnet.Context, instead of becoming Union[mxnet.Context, List[mxnet.Context]].

[haojin2]

@ChaiBapchya @nickguletskii There's one function in d2l-en that could be a possible reference.