vp9: neon: optimise convolve8_horiz functions

Each iteration of the horizontal loop reuses 7 of the 11 source
values.  Loading only the 4 new values saves some time.

Also add preload for source data.

Overall 4% faster on Chromebook.

Change-Id: I8f69e749f2b7f79e9734620dcee51dbfcd716b44