Investigate sum_3_azip's performance

[bluss]

benchmark sum_3_azip seems to perform abysmally compared with the equivalent sum_3_azip_fold, investigate why, and why the former doesn't autovectorize like the latter.

[mati865]

The difference is minimal with current master and nightly:

test sum_3_azip                        ... bench:         825 ns/iter (+/- 130)
test sum_3_azip_fold                   ... bench:         820 ns/iter (+/- 105)