[mlir] Don't iterate mutable user list

[vitalybuka]

Summary

executeOp.operandsMutable().append(asyncTokens) in
addAsyncDependencyAfter can resize and invalidate iterators.

Fixes reports like https://reviews.llvm.org/P8286

[herhut]

Thanks!

mlir/lib/Dialect/GPU/Transforms/AsyncRegionRewriter.cpp

for (size_t i = 0; i < worklist.size(); ++i) {
 
       // Erase `gpu.wait` and return async dependencies from execute op instead.
       SmallVector<Value, 4> dependencies = waitOp.asyncDependencies();
       waitOp.erase();
       executeOp = addExecuteResults(executeOp, dependencies);
 
       // Add the async dependency to each user of the `async.execute` token.
       auto asyncTokens = executeOp.getResults().take_back(dependencies.size());
-      for (Operation *user : executeOp.token().getUsers())
+      SmallVector<Operation *, 4> users(executeOp.token().user_begin(),
+                                        executeOp.token().user_end());
+      for (Operation *user : users)
         addAsyncDependencyAfter(asyncTokens, user);
     }
   }
 
 private:
   // Returns whether all token users are either 'async.execute' or 'async.await'
   // ops. This is used as a requirement for pushing 'gpu.wait' ops from a
   // 'async.execute' body to it's users. Specifically, we do not allow