[CUDA] Cholesky via cuSOLVER

[HaoXuAI]

Cholesky is a NO_GPU stub on CUDA, so mx.linalg.cholesky requires a CPU
stream. This backs it with cuSOLVER potrf / potrfBatched, and adds the
cusolver_utils handle and error plumbing that getrf, geqrf, gesvd and
syevd can reuse.

linalg::cholesky no longer forces a CPU stream when CUDA is available.
Metal behaviour is unchanged: cu::is_available() is false there, so the
existing error is still raised.

Testing

Verified on an L40S (sm_89) with CUDA 12.6, against the CPU/LAPACK path:
test_linalg.py passes 19 tests / 266 subtests, including under DEVICE=gpu.
The new test covers batched, non-contiguous and empty inputs, and asserts the
unrequested triangle is exactly zero.

Benchmarks

benchmarks/python/cholesky_bench.py, fp32, MLX_ENABLE_TF32=0, vs
torch.linalg.cholesky on the same GPU (MLX ms / PyTorch ms):

shape	MLX	PyTorch	ratio
512²	0.435	0.186	2.34x
2048²	1.891	0.951	1.99x
4096²	4.536	2.366	1.92x
4096² upper	3.830	4.332	0.88x
256 x 32²	0.089	0.058	1.53x
64 x 128²	0.223	0.134	1.66x
16 x 512² upper	0.666	0.664	1.00x

Two thirds of the single-matrix gap is potrf itself — even in its faster fill
mode 4096² takes 3.82 ms against PyTorch's 2.40 ms. The rest is that
upper=False maps to CUBLAS_FILL_MODE_UPPER, consistently slower than
FILL_MODE_LOWER. Computing the fast mode and transposing is equivalent and
measures +13% at n=4096 but -8% at n=512, so I left it out.

Note

check_cpu_stream_unless_cuda is new cross-cutting surface that each further
linalg op would touch. Happy to reshape it — a per-primitive capability query
may be preferable to a CUDA-specific branch.

[zcbenz]

Thanks for the PR! There is actually already a almost identical PR #4208 so I'm closing this as a duplicate.

[HaoXuAI]

thanks, when I started the PR was not there