Add shared Maven build workflow

[msarahan]

Why

Consolidates Java package build steps into a common workflow. Anticipated to be used for cuvs, cudf, cuopt and any other future Java packages.

Users of this script no longer configure:

• compute-matrix.yaml directly.
• strategy.matrix and fromJSON(...).
• Repeated permissions and secrets: inherit for internal jobs.
• Architecture forwarding.
• CUDA-pinned image-tag construction.
• CUDA-major deduplication.
• Per-classifier artifact-name construction.
• A second matrix definition for packaged-JAR testing.

They still own:

• Which repository script builds each Maven coordinate.
• Output file paths.
• Dependencies between coordinates, such as cuvs-lucene needing cuvs-java.
• Changed-file gating.
• Maven-repository assembly and publication.

So the downstream configuration becomes declarative: “build this coordinate using this script and image family,” instead of reconstructing the matrix execution machinery in every repository. It also centralizes the correctness-sensitive rule that a cu12 artifact must actually run in a CUDA 12-pinned image.

Deliberate publication boundary

This workflow does not gather or publish artifacts. cuOpt currently produces roughly 405 MB per classifier, so four classifiers can exceed the Sonatype Portal 1 GB bundle limit. Keeping gathering separate lets smaller artifacts use one Maven-repository-layout bundle while larger projects can split publication without changing the build contract. Maven signing and Central staging remain the responsibility of the existing maven-publish.yaml workflow.

Related: NVIDIA/cuopt#1818, rapidsai/build-infra#379

[msarahan]

testing pipeline at https://github.com/NVIDIA/cuvs/actions/runs/33665962622?pr=2545