Shuffling to memory can get out-of-sync when fetching multiple compressed map outputs

Haven't verified yet but certainly looks like tez needs same fix as MAPREDUCE-5308 in IFile.

Specifically saw this because downstream tasks were reporting enough fetch failures that long-running upstream tasks had to be re-run, which makes job run for much longer than it needs.

Usually shows itself as an "Invalid map id" error on a multi-map fetch on part 2-n (i.e. never the first one).