DMA: Increase precision for large transfers

Also gets rid of the delay on the GPU side for writing to VRAM (doesn't
make sense), and it's not needed since we slice the block transfers now.

Fixes palette corruption in Vigilante 8, and missing rider in
Championship Motocross 2001 featuring Ricky Carmichael.

src/core/gpu.h

@@ -167,7 +167,7 @@ protected:
   // The GPU internally appears to run at 2x the system clock.
   ALWAYS_INLINE static constexpr TickCount GPUTicksToSystemTicks(TickCount gpu_ticks)
   {
-    return std::max<TickCount>(gpu_ticks >> 1, 1);
+    return std::max<TickCount>((gpu_ticks + 1) >> 1, 1);
   }
   ALWAYS_INLINE static constexpr TickCount SystemTicksToGPUTicks(TickCount sysclk_ticks) { return sysclk_ticks << 1; }