Have cu12/cu13 extras depend on cuda-core extra - #457

This fixes two packaging / dependency issues currently affecting cudf's CI

1. the cudf.pandas (and pylibcudf?) oldest-dependency jobs pin `cuda-bindings==12.9.3`. A dependency on `numba-cuda-mlir` brings in a dependency on `cuda-core`, which currently resolves to `cuda-core==1.3.0` which is incompatible with `cuda-bindings<12.9.8`. NVIDIA/numba-cuda-mlir#457 attempts to fix this upstream, but that requires more work. For now, I've updated our commands that install `cuda-bindings` to *also* install `cuda-core[cu12]`. This will ensure that its `cuda-bindings` pin is honored: https://github.com/NVIDIA/cuda-python/blob/e119c3143a8cab7a2af9c600098d9c7192271d95/cuda_core/pyproject.toml#L63.
2. Similar issue for the upstream polars test environments, but this time from `torch`. Our current method of making the upstream polars test environment has some issues. 282554a has the start of a proper fix, but still needs work  to fit with how cudf-polars/polars/torch declare their dependencies.


For now, I'm comfortable with skipping the torch tests from polars' test suite. We have the `conda-python-cudf-polars-tests-integration-optional` that should still give us some coverage for that cudf-polars/torch interop.

Closes #25224.

Authors:
  - Tom Augspurger (https://github.com/TomAugspurger)
  - James Lamb (https://github.com/jameslamb)

Approvers:
  - James Lamb (https://github.com/jameslamb)

URL: #25227

This updates our package metadata to express our dependency on
cuda-core.

1. cudf and pylibucdf don't directly use cuda-core, but they do
   need a cuda-core that's compatible with the installed cuda-bindings
   thanks to a transitive cuda-core dependency via numba-cuda-mlir.
   NVIDIA/numba-cuda-mlir#457 should fix
   things upstream. In the meantime, we've added that to the `test`
   dependency
2. cudf-polars imports cuda-core, and so should express a dependency
   on it in the package metadata.

This updates our package metadata to express our dependency on cuda-core and cuda-bindings.

### cudf/pylibucdf

cudf and pylibcudf don't directly use cuda-core, but they do need a cuda-core that's compatible with the installed cuda-bindings thanks to a transitive cuda-core dependency via numba-cuda-mlir. Currently, installing numba-cuda-mlir and our pinned `cuda-bindings==12.8.3` results in an environment that breaks at runtime.

#25227 added workarounds at our install sites. @jameslamb expressed a preference to keep things centrally located in dependencies.yml. As a compromise, we've added `cuda-core[cu{12,13}]` to our `test` dependency.

NVIDIA/numba-cuda-mlir#457 (or something like it) should fix things upstream, at which point we can remove this (test) dependency.

### cudf-polars

cudf-polars imports cuda-core, and so should express a dependency on it in the package metadata.

Authors:
  - Tom Augspurger (https://github.com/TomAugspurger)
  - James Lamb (https://github.com/jameslamb)

Approvers:
  - James Lamb (https://github.com/jameslamb)
  - Bradley Dice (https://github.com/bdice)

URL: #25240