rocblas alt impl during backward pass only

[hubertlu-tw]

Cherry-picked the commit to resolve Apex compiling issues related to rocblas_gemm_flags_fp16_alt_impl implementations in MLP and MHA extensions. The error message is like

/opt/rocm-5.0.1/bin/hipcc -I/var/lib/jenkins/apex/csrc -I/opt/conda/lib/python3.7/site-packages/torch/include -I/opt/conda/lib/python3.7/site-packages/torch/include/torch/csrc/api/include -I/opt/conda/lib/python3.7/site-packages/torch/include/TH -I/opt/conda/lib/python3.7/site-packages/torch/include/THC -I/opt/conda/lib/python3.7/site-packages/torch/include/THH -I/opt/rocm-5.0.1/include -I/opt/rocm-5.0.1/miopen/include -I/opt/conda/include/python3.7m -c /var/lib/jenkins/apex/csrc/mlp_hip.hip -o build/temp.linux-x86_64-3.7/var/lib/jenkins/apex/csrc/mlp_hip.o -fPIC -D__HIP_PLATFORM_HCC__=1 -DUSE_ROCM=1 -DCUDA_HAS_FP16=1 -D__HIP_NO_HALF_OPERATORS__=1 -D__HIP_NO_HALF_CONVERSIONS__=1 -O3 -DVERSION_GE_1_1 -DVERSION_GE_1_3 -DVERSION_GE_1_5 -DTORCH_API_INCLUDE_EXTENSION_H -DPYBIND11_COMPILER_TYPE="_gcc" -DPYBIND11_STDLIB="_libstdcpp" -DPYBIND11_BUILD_ABI="_cxxabi1011" -DTORCH_EXTENSION_NAME=mlp_cuda -D_GLIBCXX_USE_CXX11_ABI=1 --amdgpu-target=gfx906 -fno-gpu-rdc -std=c++14
/var/lib/jenkins/apex/csrc/mlp_hip.hip:1518:20: error: no member named 'BackwardPassGuard' in namespace 'at'
        flag = at::BackwardPassGuard::is_backward_pass() ? rocblas_gemm_flags_fp16_alt_impl : 0;
               ~~~~^
1 error generated when compiling for gfx906.
error: command '/opt/rocm-5.0.1/bin/hipcc' failed with exit status 1

[pruthvistony]

the object in_backward, doesnt seem to be used anywhere. Please check.

[pruthvistony]

is this required?

[pruthvistony]

It is in the Node object. Please ignore previous comment.

[jithunnair-amd]

CI failed unit tests, but not related to this PR afaict. Merging to unblock QA https://ontrack-internal.amd.com/browse/SWDEV-329776

[jithunnair-amd]

Kicking off CI to get cleaner CI report (w/ updated Magma and --continue-through-error):
test pytorch please