[mlir:Async] Convert AsyncParallelFor pass to ModuleOp pass

[ezhulenev]

Summary

Depends On D104891

Outlining scf.parallel body as a function requires async-parallel-for pass to be a ModuleOp pass

[bondhugula]

Incidentally, does this pass work as intended if scf.parallel are nested within each other? I didn't see a test case where you would get an outlined function in a function already outlined.

[ezhulenev]

In theory nested scf.parallel should work, but my plan was to disable this transformation for nested operation because computing block size will get tricky. Will send a PR for this change after I'll land all the other pending changes.

I wanted too look at going directly from linalg operations to parallel loops, because in linalg it's easy to figure out the loop structure and iteration space size and compute approximate cost from it, then it should be easier to decide how many loops are worth parallelizing.

mlir/include/mlir/Dialect/Async/Passes.td

 //===-- Passes.td - Async pass definition file -------------*- tablegen -*-===//
 //
 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
 // See https://llvm.org/LICENSE.txt for license information.
 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
 //
 //===----------------------------------------------------------------------===//
 
 #ifndef MLIR_DIALECT_ASYNC_PASSES
 #define MLIR_DIALECT_ASYNC_PASSES
 
 include "mlir/Pass/PassBase.td"
 
-def AsyncParallelFor : Pass<"async-parallel-for"> {
+def AsyncParallelFor : Pass<"async-parallel-for", "ModuleOp"> {
   let summary = "Convert scf.parallel operations to multiple async compute ops "
                 "executed concurrently for non-overlapping iteration ranges";
   let constructor = "mlir::createAsyncParallelForPass()";
 
   let options = [
     Option<"asyncDispatch", "async-dispatch",
       "bool", /*default=*/"true",
       "Dispatch async compute tasks using recursive work splitting. If `false` "