... off-by one size in special cases

The last commit wasn't drastic enough.  We changed SIGNED_ROUND()
to use RINT(), which, in turn, may use rint().  However, rint()
effectively breaks ties to even, so that we get stuff like
'rint (1.5) - rint (0.5) == 2.0 - 0.0 == 2.0'.  This can't be
good--it's entirely possible that we're bitten by this in other
cases without noticing.

Avoid rint() entirely in RINT(), and always use 'floor (x) + 0.5'
instead, which always breaks ties up.  Hopefully, this doesn't
break anything else...