huge refactor for high RewritePatterns usage and less ad-hoc cpp code

remove Spatial many ops in favor of tensor ops like in pim

src/PIM/Conversion/SpatialToPim/ComputeLikeRegionUtils.hpp

@@ -0,0 +1,17 @@
+#pragma once
+
+#include "mlir/IR/Operation.h"
+#include "mlir/IR/PatternMatch.h"
+
+#include <optional>
+
+namespace onnx_mlir {
+
+std::optional<unsigned> getDirectComputeLikeInputIndex(mlir::Operation* owner, unsigned operandNumber);
+
+void replaceAndEraseDirectComputeLikeInput(mlir::PatternRewriter& rewriter,
+                                           mlir::Operation* owner,
+                                           unsigned inputIndex,
+                                           mlir::Value replacement);
+
+} // namespace onnx_mlir