General GPU/CUDA support for the JuliaManifolds ecosystem.

The package is in early stages of development, and the API is not yet stable.

Notes:

• exp! on PowerManifold(Stiefel(32, 16), 2048) is about 20x faster on CUDA.
• PolarRetraction is about 15x faster on CUDA. Batched SVD seems to work well.
• Detailed benchmarking scripts are in benchmarks/.
• QR decomposition doesn't seem to be particularly fast on GPU. Q matrix formation can't even be batched as of February 2026.