Add fallback matrix for nvcomp. (#14082)

Some platforms (such as aarch64 + CUDA 12) don't have a matching matrix entry for nvcomp. This PR adds a fallback matrix entry so it is possible to attempt local development on aarch64 with CUDA 12.

Authors:
  - Bradley Dice (https://github.com/bdice)

Approvers:
  - AJ Schmidt (https://github.com/ajschmidt8)

URL: #14082

dependencies.yaml

@@ -241,6 +241,10 @@ dependencies:
               cuda: "11.8"
             packages:
               - *nvcomp
+          # TODO: Fallback matrix for aarch64 CUDA 12. After migrating to nvcomp 3,
+          # all CUDA/arch combinations should be supported by existing packages.
+          - matrix:
+            packages:
   build_wheels:
     common:
       - output_types: pyproject