Mode switch for rempi gather

Switch to non-streaming mode to perform gather loads instead of doing
streaming-compatible gather. This requires handling different vector
lengths when passing vector parameters during the mode switch.

Adds the NOINLINE attribute to force function calls so that mode
switching happens.

Note: For the rempi functions in helpersve.h, there appears to be a bug
in gcc's optimization for passing vector lengths between modes. The
current workaround is to pass it in the output array pointer.