use libcusolver_lapack_static.a for CUDA>=12 (#98072)
Needed for https://github.com/pytorch/builder/pull/1374 to enable nightly CUDA12.1 builds.
From the cuSOLVER release notes (https://docs.nvidia.com/cuda/cusolver/index.html#link-third-party-lapack-library):
> The `liblapack_static.a` library is deprecated and will be removed in the next major release. Use the `libcusolver_lapack_static.a` instead.
Note that "next major release" corresponds to CUDA 12, not 13.
The fix was verified locally on an H100 using https://github.com/pytorch/builder/pull/1374 and pip wheels were properly built:
```
>>> torch.version.cuda
'12.1'
>>> torch.backends.cudnn.version()
8801
>>> conv =nn.Conv2d(3, 3, 3).cuda()
>>> x = torch.randn(1, 3, 224, 224).cuda()
>>> out = conv(x)
>>> out.sum()
tensor(5386.9219, device='cuda:0', grad_fn=<SumBackward0>)
```
CC @malfet @atalman @ngimel
Pull Request resolved: https://github.com/pytorch/pytorch/pull/98072
Approved by: https://github.com/malfet, https://github.com/atalman
