Use nvcomp python bindings from conda-forge instead of PyPI

[weiji14]

Available since conda-forge/nvcomp-feedstock#23

The conda-forge package actually reuses the same nvidia-nvcomp-cu12==4.2.0.14 wheel from PyPI (see https://github.com/conda-forge/nvcomp-feedstock/blob/e255600ba4eeec9fcb966b03b9c1cd76e5952af4/recipe/meta.yaml#L11-L12), so should be practically the same.

Have tested running benchmarks/zstd_benchmark.py locally, and seems to work fine:

$ python benchmarks/zstd_benchmark.py
Opened array with compressors: (ZstdCodec(level=0, checksum=False),)
Total time to read data: 1.8253955841064453 s
Effective I/O bandwidth: 0.534986776840506 GB/s
Opened array with compressors: (NvcompZstdCodec(level=0, checksum=False),)
Total time to read data: 1.3693149089813232 s
Effective I/O bandwidth: 0.7131759784361771 GB/s

[weiji14]

zarr-developers/zarr-python#2863 should be close to merging, so can also wait a few days/weeks to remove this last pip-installed package from the environment.yml.

[weiji14]

That PR is taking longer than I thought 😅, so I'll just merge this first.