rsa/rsa_ossl.c: implement variant of "Smooth CRT-RSA."

In [most common] case of p and q being of same width, it's possible to
replace CRT modulo operations with Montgomery reductions. And those are
even fixed-length Montgomery reductions...

Reviewed-by: Paul Dale <paul.dale@oracle.com>
(Merged from https://github.com/openssl/openssl/pull/6915)