Michael Weiser authored Dec 25, 2020 and Niels Möller committed Jan 13, 2021

Switch arm neon assembler routines to endianness-agnostic loads and
stores where possible to avoid modifications to the rest of the code.
This involves switching to vld1.32 for loading consecutive 32-bit words
in host endianness as well as vst1.8 for storing back to memory in
little-endian order as required by the caller. Where necessary, r3 is
used to store the precalculated offset into the source vector for the
secondary load operations. vstm is kept for little-endian platforms
because it is faster than vst1 on most ARM implementations.

vst1.x (at least on the Allwinner A20 Cortex-A7 implementation) seems to
interfer with itself on subsequent calls, slowing it down further. So we
reschedule some instructions to do stores as soon as results become
available to have some other calculations or loads before the next
vst1.x. This reliably saves two additional cycles per block on salsa20
and chacha which would otherwise be incurred.

vld1.x does not seem to suffer from this or at least not to a level
where two consecutive vld1.x run slower than an equivalent vldm.
Rescheduling them similarly did not improve performance beyond that of
vldm.

Signed-off-by: Michael Weiser <michael.weiser@gmx.de>