compact spatial IR through different new operations and dedicated syntax

fast spatial node merging with batch operations

src/PIM/Conversion/ONNXToSpatial/Patterns/NN/Pool.cpp

@@ -33,9 +33,7 @@ static int64_t getOptionalI64(std::optional<ArrayAttrT> arrayAttr, size_t index,
 
 static Value concatAlongAxis(ConversionPatternRewriter& rewriter, Location loc, int64_t axis, ArrayRef<Value> values) {
   assert(!values.empty() && "Expected at least one value to concatenate.");
-  if (values.size() == 1)
-    return values.front();
-  return tensor::ConcatOp::create(rewriter, loc, axis, values);
+  return createSpatConcat(rewriter, loc, axis, values);
 }
 
 static Value materializeContiguousTile(ConversionPatternRewriter& rewriter, Location loc, Value tile) {

src/PIM/Conversion/ONNXToSpatial/Patterns/NN/Softmax.cpp

@@ -47,8 +47,7 @@ buildSoftmax(Value input, int64_t softmaxAxis, int64_t axis, ConversionPatternRe
   for (Value slice : slices)
     rebuiltSlices.push_back(buildSoftmax(slice, softmaxAxis, axis + 1, rewriter, loc));
 
-  return rebuiltSlices.size() == 1 ? rebuiltSlices.front()
-                                   : tensor::ConcatOp::create(rewriter, loc, axis, rebuiltSlices).getResult();
+  return createSpatConcat(rewriter, loc, axis, rebuiltSlices);
 }
 
 struct SoftmaxToSpatialCompute : OpConversionPattern<ONNXSoftmaxOp> {