[XLA] Insert reduce-precision nodes inside tail of fusion nodes, and don't add duplicates.

Reduce-precision instructions can be fused, so when adding them to the output of fusion nodes, we should add them inside the node rather than outside.  In addition, since the same node may be selected by multiple passes (especially when we have fusion-by-contents passes with different trigger conditions), we want to avoid inserting redundant reduce-precision instructions.

PiperOrigin-RevId: 166526816