refactor CUDA versions in dependencies.yaml

[jameslamb]

Description

Contributes to rapidsai/build-planning#7.

Proposes splitting the cuda-version dependency in dependencies.yaml out to its own thing, separate from the bits of the CUDA Toolkit this project needs.

Benefits of this change

• prevents accidental inclusion of multiple cuda-version version in environments
• reduces update effort (via enabling more use of globs like "12.*")
• improves the chance that errors like "conda recipe is missing a dependency" are caught in CI

Checklist

• I am familiar with the Contributing Guidelines.
• New or existing tests cover these changes.
• The documentation is up to date with these changes.

[vyasr]

One suggestion, otherwise all good.

[jameslamb]

One job is failing, like this:

Collecting cudf-cu12==24.2.*,>=0.0.0a0
  Downloading https://pypi.anaconda.org/rapidsai-wheels-nightly/simple/cudf-cu12/24.2.0a177/cudf_cu12-24.2.0a177-cp310-cp310-manylinux_2_28_x86_64.whl (426.2 MB)
     ━━━━━━━━━                                97.1/426.2 MB 7.5 MB/s eta 0:00:44
ERROR: Wheel 'cudf-cu12' located at /tmp/pip-unpack-rzwlblzl/cudf_cu12-24.2.0a177-cp310-cp310-manylinux_2_28_x86_64.whl is invalid.

(build link)

To me, that looks like the wheel download was interrupted, e.g. by some brief networking disruption. Pushed an empty commit to re-trigger CI.

[jakirkham]

It looks like that cleared out

[bdice]

/merge