Cherry-pick the commit to make TORCH_(CUDABLAS|CUSOLVER)_CHECK usable in custom extensions

[hubertlu-tw]

Please see pytorch#67161 and pytorch@d4493b2 for more information. This solved the issue of import error for Apex multihead_attn extension in PyTorch 1.10.

[hubertlu-tw]

Earlier without this fix TORCH_(CUDABLAS|CUSOLVER)_CHECK is not usable in custom extensions. For instance, while using/importing Apex Multihead Attention module (from apex.contrib.multihead_attn import EncdecMultiheadAttn) it throws some error messages as follows:
from apex.contrib.multihead_attn import EncdecMultiheadAttn /opt/conda/lib/python3.8/site-packages/apex/contrib/multihead_attn/__init__.py:1: in <module> from .self_multihead_attn import SelfMultiheadAttn /opt/conda/lib/python3.8/site-packages/apex/contrib/multihead_attn/self_multihead_attn.py:9: in <module> from .fast_self_multihead_attn_func import fast_self_attn_func /opt/conda/lib/python3.8/site-packages/apex/contrib/multihead_attn/fast_self_multihead_attn_func.py:2: in <module> import fast_self_multihead_attn E ImportError: /opt/conda/lib/python3.8/site-packages/fast_self_multihead_attn.cpython-38-x86_64-linux-gnu.so: undefined symbol: _ZN2at4cuda4blas19_cublasGetErrorEnumE15rocblas_status_. The reason is because TORCH_CUDABLAS_CHECK is extensively used in Apex Multihead Attention module.

Docker: rocm/pytorch:rocm4.5.2_ubuntu18.04_py3.8_pytorch_1.10.0 (the current rocm/pytorch:latest)
Command to reproduce the error: from apex.contrib.multihead_attn import EncdecMultiheadAttn

[jithunnair-amd]

This cherry-pick is only needed for the release/1.10 branch as the corresponding PyTorch commit is already present in the rocm5.0 and rocm5.1_internal_testing branches.