Add cuSPARSELt to the nightlies

[jcaip]

Issue description

cuSPASRELt is a library that offers accelerated semi-structured sparse matmul. We are looking to integrate v0.5.0
The download link can be found here: https://developer.nvidia.com/cusparselt-downloads
The library is ~ 20 Mb.

We merged build support for cuSPASRELt in this PR
To build pytorch with cuSPASRELt, we need to set two env variables,

USE_CUSPARSELT=1
CUSPARSELT_ROOT=/path/to/cusparselt/download

cc @seemethere @malfet

[jcaip]

Hi just FYI the cusparselt version has been bumped to v0.5.0, we'd like to use the newest version

[malfet]

Updates:

• 0.5.0 binary size is 40Mb rather than 20 as claimed in issue description
• Added it to the binary docker images in pytorch/builder@00841b6
• It turns out that library requires a much too modern libc/libstdc++, see https://github.com/pytorch/pytorch/actions/runs/6478662398/job/17593042842#step:15:15786 for example

/opt/rh/devtoolset-9/root/usr/libexec/gcc/x86_64-redhat-linux/9/ld: /usr/local/cuda/lib64/libcusparseLt.so.0: undefined reference to `std::__cxx11::basic_string<char, std::char_traits<char>, std::allocator<char> >::_M_construct(unsigned long, char)@GLIBCXX_3.4.21'

[malfet]

Ok, but cuSparseLT-0.4.0 seems to be compatible with CentOS7, let's try that...

[malfet]

Added to nightly wheels, not sure if/when we need it for conda, but closing for now