split up CUDA-suffixed dependencies in dependencies.yaml

[jameslamb]

Description

Contributes to rapidsai/build-planning#31

Follow-up to #15245

RAPIDS DLFW builds prefer to build all RAPIDS packages together without CUDA suffixes, leading to the following set of requirements for cudf wheels built there:

• project name must be cudf (not cudf-cu12)
• all dependencies must be unsuffixed (e.g. rmm not rmm-cu12)
• the correct set of dependencies based on CUDA version must be expressed in the wheel metadata (e.g. cubinlinker and ptxcompiler on CUDA 11, pynvjitlink on CUDA 12)

To meet all 3 of those, this proposes decomposing CUDA-suffixed dependencies in dependencies.yaml into two lists... cuda_suffixed="true" and cuda_suffixed="false".

That'd allow DLFW builds to do the following to meet its requirements:

pip wheel \
  -C rapidsai.disable-cuda=true \
  -C rapidsai.matrix-entry="cuda=12.5;cuda_suffixed=false" \
  .

Checklist

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

Notes for Reviewers

Why still have the fallback matrices be CUDA 11 dependencies?

I'd originally included changese for rapidsai/build-planning#68 here, but chose to revert them to reduce the size of this diff and to make it absolutely clear... the fallback matrices' values are not relevant for devcontainers, DLFW, or normal wheel builds.