Share gfx12 K enc0 V tiles while preserving strict FP16 arithmetic

kernels/k/dltss_pwin_enc0_layer.hip

@@ -1,5 +1,19 @@
 // CU mode: the waves of a window share one L0 cache (1-6% faster here; slower in the wide layers)
 // d4r-build-flags: -mcumode
 #define PWIN_F32ACC
+#ifndef D4R_K_POS_MT
+#define D4R_K_POS_MT 2
+#elif !defined(__GFX12__) || !defined(D4R_K_FP16_BASELINE)
+#error "experimental position token tiling requires gfx12 strict K baseline"
+#endif
+// Reuse V across the window's waves. Paired gfx1201 replays retain every half
+// code and reduce enc0 time by about 5%; gfx11 and arithmetic experiments keep
+// their previous path. D4R_K_PRIVATE_V supplies the original A/B control.
+#if defined(__GFX12__) && defined(D4R_K_FP16_BASELINE) && !defined(D4R_K_PRIVATE_V) && \
+    !defined(D4R_K_PACKED_ACC) && !defined(D4R_K_F16_WMMA)
+#ifndef POS_VSHARE
+#define POS_VSHARE
+#endif
+#endif
 #include "pwin_pos.h"
-POS_ENCODER(dltss_pwin_enc0_layer, 2, 32, 64, 16, 2)
+POS_ENCODER(dltss_pwin_enc0_layer, 2, 32, 64, 16, D4R_K_POS_MT)