[ET-VK][reduction] Work around Adreno 8x4 dispatch failure

## Ulterior Motive

Restore ET-VK operator correctness on Galaxy S25 devices used by Eureka CI.

## Rationale

**What**: Make `reduce_nworkers()` select four workers when scaling would
select eight.

**Why**: Galaxy S25's Adreno Vulkan driver skips alternating work groups for
the resulting `8x4` local size. Reduction outputs retain uninitialized data
for those groups.

## Details

| Behavior | Before | After |
| --- | --- | --- |
| Short reductions selecting eight workers | Dispatch `8x4`; eight generated tests fail | Dispatch `4x4`; all generated tests pass |
| Larger reductions | Scale to 16, 32, or 64 workers | Unchanged |

Authored with Codex.

Differential Revision: [D122891447](https://our.internmc.facebook.com/intern/diff/D122891447/)


ghstack-source-id: 440132327
Pull-Request: pytorch#23325

Author: SS-JIA

backends/vulkan/runtime/graph/ops/impl/Reduce.cpp

@@ -119,7 +119,9 @@ uint32_t reduce_nworkers(
   while (nworkers * 2u <= cap && nworkers < extent) {
     nworkers *= 2u;
   }
-  return nworkers;
+  // Some Adreno drivers skip alternating work groups for an 8x4 local size.
+  // Keep short reductions on the portable four-worker geometry.
+  return nworkers == 8u ? 4u : nworkers;
 }
 
 GlobalWorkGrid reduce_gwg_impl(