[mlir][linalg][bufferize][NFC] scf.execute_region result considered a write

[springerm]

Summary

This change by itself it not very useful because bufferization of scf.execute_region is not yet implemented, but it avoids an edge case in a subsequent commit related to partial bufferization.

Depends On D114616

[springerm]

Updated comment and squashed into dependent commit. The main problem here was an inconsistent op interface implementation.

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

struct ExecuteRegionOpInterface
 
   bool mustBufferizeInPlace(Operation *op, OpResult opResult) const {
     // ExecuteRegionOp results always bufferize in-place. Since they have no
     // OpOperands, they are mostly ignored by the analysis once alias sets are
     // set up.
     return true;
   }
 
+  // TODO: For better bufferization results, this could return `true` only if
+  // there is a memory write in the region.
+  bool isMemoryWrite(Operation *op, OpResult opResult) const {
+    // Similar to scf.if, results of this op are always considered memory writes
+    // in the analysis.
+    // TODO: It seems that "isMemoryWrite = true" is a useful pattern for all
+    // ops that may yield values that are captured from outside of the op
+    // (as opposed to BlockArguments of the op). This should be further explored
+    // in the design doc.
+    return true;
+  }
+
   LogicalResult bufferize(Operation *op, OpBuilder &b,
                           BufferizationState &state) const {
     // TODO: Add bufferization support when needed. scf.execute_region should be
     // bufferized similar to scf.if.
     auto executeRegionOp = cast<scf::ExecuteRegionOp>(op);
     bool hasTensorReturnType = any_of(
         op->getResultTypes(), [](Type t) { return t.isa<TensorType>(); });
     if (hasTensorReturnType)