Improve cc version detection for cublaslt

[BlackHC]

The current code fails to detect cublaslt for new cc versions (e.g. 8.9 fails the check).

This fixes issue #97.

[TimDettmers]

Thank you for this contribution!

[gururise]

I don't think this PR quite fixes issue #97, which referenced lack of Int8 Matmul error being thrown on the Nvidia Tesla P40.

Still getting the following error:

=============================================
ERROR: Your GPU does not support Int8 Matmul!
=============================================

python: /home/gene/dockerx/temp/bitsandbytes/csrc/ops.cu:408: int igemmlt(cublasLtHandle_t, int, int, int, const int8_t*, const int8_t*, void*, float*, int, int, int) [with int FORMATB = 3; int DTYPE_OUT = 32; int SCALE_ROWS = 0; cublasLtHandle_t = cublasLtContext*; int8_t = signed char]: Assertion `false' failed.

I created a new issue #118 to make it clear that problem still exists.

[BlackHC]

@gururise

The following has helped me better understand what was going on, but your error sounds very different to the one I fixed with my PR.

logging.basicConfig(level=logging.DEBUG)
from bitsandbytes.cextension import CUDASetup
CUDASetup.get_instance().print_log_stack()

Best wishes,
Andreas