The Adler32 digest consists of two 16-bit words whose values are
calculated modulo 65521 (largest prime < 2^16). To avoid two division
instructions per byte, this version copies an optimization found in
zlib which defers the modulus until close to the point that the
intermediate sums can overflow 2^32. (zlib uses NMAX == 5552 for
this, this version uses 5000)

The bug is that in the deferred modulus case, the modulus was
only applied to the high word (and twice at that) but not to
the low word. The fix is to apply it to both words.

Reported by: Miod Vallat <miod@openbsd.org>