WIP: Expand accumlator data type precision for reductions to address accuracy issues

The MR proposes changes to expand the reduction accumulator type to something other than the input type to address accuracy issues that were encountered with bfloat16 reductions.

These changes were originally part of the bfloat16 for HIP MR but were separated out into this MR for clarity.

This proposal is a work in progress and will need close scrutiny.

/cc @cantonios