In order to support flash attention, we'd need to be able to expand ReduceOps in the reduction dimension as well. We will do this by expanding the source of ReduceOp and locally reduce all of them. In that effort, we introduce this PR(1st out of 2) that add support of locally reducing over multiple variables.

The second PR on the way would be expansion of ReduceOp.

In this PR we are contributing two things:

1. Checks for consistency of indexing_dims, types, thread_shapes for multiple sources of ReduceOp
2. Modify emit of local reduction to generate iteratively slice and reduce over multiple arguments/srcs.