[Stream] Teach allocation of encoding to take bcast map into account.

[hanhanW]

The bcast_map represents the iteration domain mapping from the source of dispatch to matmul. So it uses bcast_map in getMapForOperandIndex() when it is present.

[hanhanW]

@benvanik you can ignore the PR because it is not merging into main. We're using a remote branch to brainstorm what's needed in encodings.

[hanhanW]

@Max191 I'm thinking that we can just replace the corresponding indexing_map with the bcast_map now.. We can look at details after we have an e2e example.

[hanhanW]

oh, there is a bug in the indexing map, fixing...

[hanhanW]

I'm thinking that we can just replace the corresponding indexing_map with the bcast_map now.. We can look at details after we have an e2e example.

So I think we only need the composed indexing_map, see the changes in the PR.