[rocm-libraries] ROCm/rocm-libraries#4652 (commit 39a5a53)

Revert "[CK] Add new fwd conv fp16/bf16 instances optimized for unit group size." (#4652)

PR ROCm/rocm-libraries#4275 contains CK fwd conv instances optimized for
`gfx950` and they do not compile for other architectures such as
`gfx940`. To ensure that the optimized instances are compiled only for
`gfx950`, compile-time guard `#if defined(CK_USE_GFX950)` was used. This
approach works correctly when we compile for a single architecture, but
when we compile simultaneously for multiple architectures, flag
`CK_USE_GFX950` is set for non-gfx950 archs as well. As a result, the
multi-arch compilation fails. The problem doesn't appear in the ROCm
libraries CI/CD pipeline since only one architecture is compiled at a
time. Hence, the CI/CD passed for the original PR.

Co-authored-by: Illia Silin <98187287+illsilin@users.noreply.github.com>