[mlir][linalg] ComprehensiveBufferize: Do not copy InitTensorOp results

[springerm]

Summary

E.g.:

%2 = memref.alloc() {alignment = 128 : i64} : memref<256x256xf32>
%3 = memref.alloc() {alignment = 128 : i64} : memref<256x256xf32>

// ... (%3 is not written to)

linalg.copy(%3, %2) : memref<256x256xf32>, memref<256x256xf32>
vector.transfer_write %11, %2[%c0, %c0] {in_bounds = [true, true]} : vector<256x256xf32>, memref<256x256xf32>

Avoid copies of %3 if %3 came directly from an InitTensorOp.

Depends On D109741

mlir/lib/Dialect/Linalg/Transforms/ComprehensiveBufferize.cpp

static LogicalResult bufferize(OpBuilder &b, VectorTransferOpInterface op,
 
   // If transfer_write is not inPlace, allocate a new buffer.
   Value newInputBuffer;
   if (inPlace != InPlaceSpec::True) {
     // Alloc a copy for `writeOp.source()`, it will become the result buffer.
     newInputBuffer = createNewAllocDeallocPairForShapedValue(
         b, loc, writeOp.source(), aliasInfo);
     Value v = lookup(bvm, writeOp.source());
+    if (!isInitTensorOp(writeOp.source()))
       b.create<CopyOp>(loc, v, newInputBuffer);
   } else {
     // InPlace write will result in memref.tensor_load(x) which must
     // canonicalize away with one of it uses.
     newInputBuffer = lookup(bvm, writeOp.source());
     assert(newInputBuffer && "missing buffer");
   }
 
   // Create a new transfer_write on buffer that doesn't have a return value.