After all the worry about accuracy in the polynomial expansion,
I figured it was time to actually add tests verifying the error
bounds here in a more reasonable way. The limits are set sort-of
randomly, more-or-less to what the new code already handles,
rounded up a bit. (The old texture-based code was _way_ worse than this,
it seems, probably due to texel center issues.)
This also shows that you can probably do 10-bit processing with
Movit without losing way too much accuracy, but that 12-bit is
too much for fp16 to handle.