[ROCm] Improve reduction sum performance

[jerrymannil]

• Use input vectorization for reduction_on_fastest_striding_dimension when dim0 >= 128

Reproducer:

import time
import torch

shapes = [
    (5079670, 128)
]

dims = [
    (1)
]

for i, shape in enumerate(shapes):
    x = torch.randn(shape, device='cuda', dtype=torch.float)
    for _ in range(10):
        w = torch.sum(x, dims[i])
    torch.cuda.synchronize()
    print(w.size())

    start_time = time.time()
    for _ in range(50):
        _ = torch.sum(x, dims[i])
    torch.cuda.synchronize()
    end_time = time.time()
    mean_time = (end_time - start_time)/50
    print(f"Avg time for shape {shape}: {mean_time * 1e6:.2f} us")

Before (MI300X):
Avg time for shape (5079670, 128): 1629.99 us

After (MI300X)
Avg time for shape (5079670, 128): 1008.59 us

cherry-pick of pytorch#160466

Fixes SWDEV-546136

Cherry-picked to release/2.8 branch via #2505

[rocm-repo-management-api]

Jenkins build for 18528f45f38a0cb0eab9c869f1c367156a1d7122 commit finished as FAILURE
Links: Blue Ocean view / Build artifacts

[pruthvistony]

Please cherry-pick into all required branches.

[jerrymannil]

! cherry-pick --onto release/2.8

[jithunnair-amd]

! cherry-pick --onto release/2.8

[dhonnappa-amd]

Created branch autogenerated/release/2.8_cherry-pick_pr-2492 and #2505

[pruthvistony]

! cherry-pick --onto rocm7.1_internal_testing