cherry-pick of pytorch#160979
Less-performant fix until pytorch#161180
is finalized

* The global reduction path in reduction kernel currently has two
threadfence operation
* The first threadfence is executed by all threads in all the blocks,
whereas the second threadfence is only run by threads in a single block
* For AMD gpus, threadfence is a heavy weight operation, esp. when run
by all the threads in the system (due to cross-XCD synchronizations)
* So using fine-grain fence gives significant performance boost for AMD
gpus.
* We do a release fence when threads write to reduce buffer in global
memory; and then do a acquire fence when threads read from the reduce
buffer

Co-author: @amd-hhashemi, @jeffdaily 

**Reproducer**:
```import time
import torch

shapes = [(2, 896, 59, 91),
]

dims = [(2, 3),
]

for i, shape in enumerate(shapes):
    x = torch.randn(shape, device='cuda', dtype=torch.bfloat16)
    x = x.to(memory_format=torch.channels_last)
    for _ in range(20):
        _ = torch.sum(x, dims[i], keepdim=True, dtype=torch.bfloat16)
    torch.cuda.synchronize()

    start_evt = torch.cuda.Event(enable_timing=True)
    end_evt = torch.cuda.Event(enable_timing=True)
    start_evt.record()
    for _ in range(100):
        _ = torch.sum(x, dims[i], keepdim=True, dtype=torch.bfloat16)
    end_evt.record()
    torch.cuda.synchronize()
    print(f"Avg time for shape {shape}: {start_evt.elapsed_time(end_evt) / 100 * 1e3:.2f} us")
```

Fixes SWDEV-545710