[SYCL][Reduction] Limit reduction work non-uniform case

[steffenlarsen]

In certain cases non-uniform reductions will not have enough work to fill the partial reduction in the last work-group. Previously we would pad with the identity to work around this, but to avoid needing to do this we can now instead reduce the size of the partial reduction to the actual amount of work.

[steffenlarsen]

HIP Timed Out Tests (4):
SYCL :: KernelAndProgram/spec_constants_after_link.cpp
SYCL :: Regression/kernel_bundle_ignore_sycl_external.cpp
SYCL :: Regression/same_unnamed_kernels.cpp
SYCL :: Regression/static-buffer-dtor.cpp
These are currently not supported on HIP and are marked unsupported in intel/llvm-test-suite#1639.

HIP Unresolved Tests (1):
SYCL :: AsyncHandler/default_async_handler.cpp
Reported in #8555.

[dm-vodopyanov]

Post-commit failure:

• Linux + Clang + Shared libs https://github.com/intel/llvm/actions/runs/4353427009/jobs/7607453013
• macOS / Build https://github.com/intel/llvm/actions/runs/4353427009/jobs/7607453421

[steffenlarsen]

Post-commit failure:

• Linux + Clang + Shared libs https://github.com/intel/llvm/actions/runs/4353427009/jobs/7607453013
• macOS / Build https://github.com/intel/llvm/actions/runs/4353427009/jobs/7607453421

Fixed in #8558.