[BACKEND] Support MMA V3 with register operand (#2375)

MMA V3 support taking operand A from register. This helps for chained
matmul operations like in attention.
Add an optimization to use this mode when it helps and add the lowering
for it.

lib/Analysis/Allocation.cpp

@@ -96,6 +96,8 @@ SmallVector<unsigned>
 getScratchConfigForCvtLayout(triton::gpu::ConvertLayoutOp op, unsigned &inVec,
                              unsigned &outVec) {
   auto repShape = getRepShapeForCvtLayout(op);
+  if (repShape.empty())
+    return repShape;
 
   auto srcTy = op.getSrc().getType().cast<RankedTensorType>();
   auto dstTy = op.getResult().getType().cast<RankedTensorType>();