[BACKEND] Fix scan issues on repetitive warps and improve perf when there's a single warp on the axis (#2330)

1. On the axis, using `getAxisNumWarpsWithUniqueData` instead of getting
the raw number of warps to avoid communication among warps that handle
the same piece of data.
2. When there's a single warp on the axis, using warp Intrinsics for
communication and skip shared memory.

Need a follow up PR for code clean up.

include/triton/Analysis/Utility.h

@@ -88,6 +88,8 @@ public:
   unsigned getNonAxisNumThreadsPerCTA();
   // Return the number of warps per CTA along axis dim.
   unsigned getAxisNumWarps();
+  // Return the number of warps per CTA along axis dim with unique data.
+  unsigned getAxisNumWarpsWithUniqueData();
   // Return the number of threads per warp along axis dim.
   unsigned getAxisNumThreadsPerWarp();
   // Return the number of blocks along axis dim.