Pass Cholesky workspace as buffer_view

[KerelosGayed]

Follow-up to #11.

cusolver_potrf took the workspace as a separate pointer and size, which is how the size()/count() mismatch in #11 slipped in. It now takes an xpu::buffer_view<T>, so the pointer and element count come from the same object, and the call site passes workspace_.view().

To support this, buffer_view gets a data() accessor that follows the same constness deduction as operator[].

This is the only internal helper that passes a pointer with a separate length. cusolver_getrf takes no workspace size, and fill_n/copy_n are left alone since they are public std-style APIs.

Validation:

• Built and ran the CPU buffer tests, including new checks for buffer_view::data() on mutable, const, empty, and owner views.
• Syntax-checked the CPU linear algebra tests.
• git diff --check passed.
• CUDA compilation/runtime tests were not run locally because nvcc is unavailable.

[karl-kes]

Thanks for fixing!