fix: do not require packaging for Torch 2.3.0+ datatype support

[akx]

What does this PR do?

This PR drops the dependency on packaging; looking at whether torch.uint64 exists is enough of a canary to see whether Torch 2.3.0 is in use and the new uint types are there.

$ uvx --python=3.10 --with=torch==2.0.0 python -c 'import torch; print(torch.__version__, hasattr(torch, "uint16"), hasattr(torch, "uint32"), hasattr(torch, "uint64"))'
2.0.0 False False False
$ uvx --python=3.10 --with=torch==2.1.0 python -c 'import torch; print(torch.__version__, hasattr(torch, "uint16"), hasattr(torch, "uint32"), hasattr(torch, "uint64"))'
2.1.0 False False False
$ uvx --python=3.10 --with=torch==2.2.0 python -c 'import torch; print(torch.__version__, hasattr(torch, "uint16"), hasattr(torch, "uint32"), hasattr(torch, "uint64"))'
2.2.0 False False False
$ uvx --python=3.10 --with=torch==2.3.0 python -c 'import torch; print(torch.__version__, hasattr(torch, "uint16"), hasattr(torch, "uint32"), hasattr(torch, "uint64"))'
2.3.0 True True True

Beyond the unnecessary dependency, this is also faster and that's not counting the import time for packaging...

$ uvx --python=3.10 --with=torch==2.3.0 --with=packaging python -m timeit -s 'from packaging.version import Version; import torch' 'Version(torch.__version__) >= Version("2.3.0")'
100000 loops, best of 5: 3.72 usec per loop
$ uvx --python=3.10 --with=torch==2.3.0 --with=packaging python -m timeit -s 'import torch' 'hasattr(torch, "uint64")'
10000000 loops, best of 5: 30.6 nsec per loop

Follows up on #664, #643.

[akx]

Rebased post #710. Strictly speaking the hasattr checks aren't required anymore since the minimum bound is torch 2.4.0, but there's a whole bunch of other torch version check stuff that became obsolete with #710 still hanging around, so another PR would be good to clean that up.