diag: disable O rescale properly, test n=128+256 baseline

tests/unit/test_fmha_v3_stage_c.py

@@ -375,10 +375,8 @@ class FmhaV3StageCMulti:
                 cute.arch.fence_view_async_tmem_store()
 
                 # === Per-tile O rescale: O *= acc_scale for kt > 0 ===
-                # TEMP: disabled for diagnosis — O rescale not applied
-                if kt > 0:
-                    pass  # O rescale disabled
-                if kt > 0:
+                # DIAG: O rescale DISABLED — testing baseline without rescale
+                if False:
                     for i in range(n_corr_tiles):
                         tTMEM_LOADtO_i = cute.make_tensor(
                             tTMEM_LOADtO.iterator + i * corr_tile_size,