vulkan: restore -inf check in FA shaders (#19582)

ggml/src/ggml-vulkan/vulkan-shaders/flash_attn_cm1.comp

@@ -42,6 +42,8 @@ D_TYPE perElemOpGqaStore(const in uint32_t r, const in uint32_t c, const in D_TY
     return elem;
 }
 
+shared float tmpsh[row_split];
+
 const uint32_t qstride = HSK_pad / 4 + 2; // in units of f16vec4
 shared f16vec4 Qf[Br * qstride];
 
@@ -213,6 +215,19 @@ void main() {
                         }
                     }
                 }
+                // skip the block if the mask is entirely -inf
+                bool all_less = subgroupAll(max_mask <= NEG_FLT_MAX_OVER_2);
+                barrier();
+                if (gl_SubgroupInvocationID == 0) {
+                    tmpsh[gl_SubgroupID] = all_less ? NEG_FLT_MAX_OVER_2 : 0.0f;
+                }
+                barrier();
+                [[unroll]] for (uint s = 0; s < gl_NumSubgroups; ++s) {
+                    max_mask = max(max_mask, tmpsh[s]);
+                }
+                if (max_mask <= NEG_FLT_MAX_OVER_2) {
+                    continue;
+                }
             }
         }