[mlir][linalg] Fix windows build issue in hoist padding.

[gysit]

Summary

Iterating backwardSlice and removing elements at the same time can fail on windows for specific build configurations (the code was introduced in https://reviews.llvm.org/D114420). This revision introduces a second vector to collect all operations and removes them after finishing the reverse iteration.

[gysit]

@hpmorgan I hope I found the right henry morgan and this fixes the issue. Let's follow up here!

[hpmorgan]

Looks good to me. Thanks Tobias!

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

HoistingAnalysis::dropNonIndexDependencies(PadTensorOp padTensorOp,
   //       %ubi = affine.min #map(%i)
   //       %ubj = affine.min #map(%j)
   //       %slice = tensor.extract_slice %source [%i, %j] [%ubi, %ubj]
   //       %padded_slice = linalg.pad_tensor %slice
   // ```
   // After iterating `backwardSlice` we obtain:
   // indexEdges = [%i, %j, %ubi, %ubj]
   // backwardSlice = backwardSlice / [linalg.fill(%cst, %arg1), scf.for %k]
+  SetVector<Operation *> operationsToRemove;
   for (Operation *op : llvm::reverse(backwardSlice)) {
     // Add the index operands of `padTensorOp` and `sliceOp` to start the
     // exploration of the index computation.
     if (op == padTensorOp || op == sliceOp) {
       addIndexOperandsToIndexEdges(op);
       continue;
     }
     // Add the index operands of the loop if its induction variable is
if (hasIndexResult(op)) {
       bool hasMemoryEffect = effectInterface && !effectInterface.hasNoEffect();
       if (hasMemoryEffect || op->getNumRegions() != 0) {
         LLVM_DEBUG(DBGS() << "Unsupported op with region or memory effect: "
                           << op << " -> skip\n");
         return failure();
       }
       continue;
     }
-    // Remove all other operation not used by the index computation except for
-    // constant operations that may be padding values used by `padTensorOp`.
+    // Remove all other operations not used by the index computation. An
+    // exception are constant operations that may be used by `padTensorOp`.
     if (!isa<arith::ConstantOp>(op))
-      backwardSlice.remove(op);
+      operationsToRemove.insert(op);
   }
+  backwardSlice.set_subtract(operationsToRemove);
   return success();
 }
 
 SmallVector<Value>
 HoistingAnalysis::getPackedTensorSizes(ImplicitLocOpBuilder &b) {
   SmallVector<Value> dynamicTensorSizes;
 
   // Upper bound the packing loop lengths to size the packed tensor. Taking