The BIGNUM behaviour is supposed to be "consistent" when going into and
out of APIs, where "consistent" means 'top' is set minimally and that
'neg' (negative) is not set if the BIGNUM is zero (which is iff 'top' is
zero, due to the previous point).

The BN_DEBUG testing (make test) caught the cases that this patch
corrects.

Note, bn_correct_top() could have been used instead, but that is intended
for where 'top' is expected to (sometimes) require adjustment after direct
word-array manipulation, and so is heavier-weight. Here, we are just
catching the negative-zero case, so we test and correct for that
explicitly, in-place.

Change-Id: Iddefbd3c28a13d935648932beebcc765d5b85ae7
Signed-off-by: NGeoff Thorpe <geoff@openssl.org>
Reviewed-by: NRichard Levitte <levitte@openssl.org>
(Merged from https://github.com/openssl/openssl/pull/1672)