Add xpu support to decoder benchmark scripts

[jonrecker]

• Add check_decoder_support function to benchmark_decoders_library
• Add argument parsing in generate_readme_data:
  --decoders option to specify list of decoders to test
  --resize_device option to specify device for resize
• Warn and skip tests if requested backends (xpu, cuda) or implementations (torchaudio, torchvision) are not supported

New options for generate_readme_data.py:

python3 generate_readme_data.py --help
usage: generate_readme_data.py [-h] [--decoders DECODERS] [--resize_device RESIZE_DEVICE]

options:
  -h, --help            show this help message and exit
  --decoders DECODERS   Comma-separated list of decoders to benchmark. Valid options: cpu, cuda, xpu, torchvision, torchaudio
  --resize_device RESIZE_DEVICE
                        Device for resize. Default: cuda if available, else cpu. Valid options: cpu, cuda, xpu

Sample command-lines to run benchmark scripts with xpu enabled:

python3 benchmark_decoders.py --decoders torchcodec_core:device=xpu,torchcodec_core_batch:device=xpu,torchcodec_public:device=xpu,torchcodec_public_nonbatch:device=xpu

python3 gpu_benchmark.py --devices xpu,cpu --resize_devices xpu,cpu,native,none

python3 generate_readme_data.py --decoders cpu,xpu

[dvrogozh]

This does not quite help:

$ python3 benchmark_decoders.py --decoders decord,decord_batch,torchcodec_core,torchcodec_core_batch,torchcodec_core_compiled,torchcodec_public,torchcodec_public_nonbatch
Traceback (most recent call last):
  File "/home/dvrogozh/git/pytorch/torchcodec/benchmarks/decoders/benchmark_decoders.py", line 172, in <module>
    main()
  File "/home/dvrogozh/git/pytorch/torchcodec/benchmarks/decoders/benchmark_decoders.py", line 129, in main
    decoders_to_run[display] = kind(**options)
  File "/home/dvrogozh/git/pytorch/torchcodec/benchmarks/decoders/benchmark_decoders_library.py", line 382, in __init__
    from torchvision.transforms import v2 as transforms_v2
ModuleNotFoundError: No module named 'torchvision'

[jonrecker]

This check was only enabled in generate_readme_data.py since the other scripts allow specifying which decoders to test. But it is no problem to add checks to the other benchmarks also. Latest commit updates benchmark_decoders.py and gpu_benchmark.py to fail gracefully if dependencies are missing, and run the remaining tests if possible.