use regs for multi reduction per block to avoid smem race

[liqiangxl]

when total_reduction_numel <= 1024, scheduler may use multiple reductions per block with bdimy > 1, this leads to race condition in shared memory when using async copy. Adding cp.async.wait_allafter the 1st async copy can avoid the race, but needs to figure out the root cause before we can safely use it. So, here we put all buffers in registers.

race detected with:

NVFUSER_DUMP=scheduler_params,cuda_to_file NVFUSER_ENABLE=kernel_debug PYTORCH_NO_CUDA_MEMORY_CACHING=1 compute-sanitizer --tool racecheck --racecheck-detect-level info  ./nvfuser_tests --gtest_filter='CombinedSchedulerTest.LayerNormBackward/dtype_double_batch_216_hidden_96'

[liqiangxl]

!test