Fix CUDA Cholesky workspace handling

[KerelosGayed]

CUDA Cholesky factorization called buffer::size(), which does not exist, preventing the float and double CUDA factorization paths from compiling.

Replace workspace_.size() with workspace_.count() at the call site to pass the requested workspace element count to the existing cusolver_potrf helper. Its signature stays unchanged.

Validation:

• git diff --check passed.
• Inspected the existing float/double CUDA Cholesky tests, which exercise this call through factorization and solve.
• CUDA compilation/runtime tests were not run locally because nvcc is unavailable.

[karl-kes]

Could you keep cusolver_potrf's signature as it was and just swap workspace_.size() for workspace_.count() at the call site? The next step would be to use buffer_view<T> where applicable in a follow up PR

[karl-kes]

LGTM