feat(ci): add CUDA 13 wheel builds

[abetlen]

Summary

• add CUDA 13.0.2 and 13.2.1 to the CUDA wheel build matrix
• document cu130 and cu132 prebuilt wheel index tags
• use a CUDA 13-specific architecture list because CUDA 13 no longer supports pre-Turing offline targets
• add an Unreleased changelog entry

Details

CUDA 13 NVCC recognizes sm_75 and newer GPU code targets, so the existing CUDA 12 architecture list cannot be reused as-is because it includes Pascal and Volta targets. CUDA 12 wheels keep the current sm_60+ coverage, while CUDA 13 wheels build sm_75+ with the existing forward-compatible PTX target.