Improve the way we allocate and use memory for GPU batch decoding

[ahmadsharif1]

Currently when we are doing batch decoding, using get frames_at_indexes, we allocate the memory on the host, and then we all allocate each frame independently and then copy the memory into the batch memory.

Memory is allocated here for the batch:

torchcodec/src/torchcodec/decoders/_core/VideoDecoder.cpp

Line 1021 in f4065f1

BatchDecodedOutput output(frameIndexes.size(), options, streamMetadata);

The copy is done here from frame to batch memory:

torchcodec/src/torchcodec/decoders/_core/VideoDecoder.cpp

Line 1031 in f4065f1

output.frames[i++] = frame;

This is wasteful especially when we are doing GPU recording because we incur multiple device to host transfers -- one per frame.

Action items:

1. We should respect the device when we are allocating the batch tensor memory.
2. We should directly use the batch sensor memory instead of incurring an extra memcpy.

[NicolasHug]

Adding bug label since batch-decoding on CUDA currently returns a CPU tensor (ref: #268)

[NicolasHug]

This was incorrectly closed by #266, I think this is still outstanding

[ahmadsharif1]

Is this still an issue? I thought I fixed this already in #319

[NicolasHug]

Yes, this was fixed in #319 (which wasn't landed yet when I re-opened this :) )