improve handling at end of buffer

a prior change reduced iterations through the input buffer to avoid the
NEON operations from overrunning the end of the locally allocated
buffer. While avoiding the overrun, it generated bad results.
Here we instead extend the locally allocated buffers enough that the
original iteration count won't overrun.

Some pre-existing bit-exact issues remain.

Bug: 136616344
Test: CTS + bit-exact cross-checks.
(cherry picked from commit aae866aed579da4e1c3299a1e9b94a1713a0decb)

Merged-In: Ifb790a7d1d09a4ce7da900b43e3fa1f7ab01ac53
Change-Id: Ia4a94c89979d6b3b0ddead135036aec40259f53a