Previously, the verification was only done when using the CRT method,
as the CRT method has been shown to be extremely sensitive to fault
attacks. However, there's no reason to avoid doing the verification
when the non-CRT method is used (performance-sensitive applications
should always be using the CRT-capable keys).

Previously, when we detected a fault (attack) through this verification,
libcrypto would fall back to the non-CRT method and assume that the
non-CRT method would give a correct result, despite having just
detecting corruption that is likely from an attack. Instead, just give
up, like NSS does.

Previously, the code tried to handle the case where the input was not
reduced mod rsa->n. This is (was) not possible, so avoid trying to
handle that. This simplifies the equality check and lets us use
|CRYPTO_memcmp|.

Change-Id: I78d1e55520a1c8c280cae2b7256e12ff6290507d
Reviewed-on: https://boringssl-review.googlesource.com/7582
Reviewed-by: David Benjamin <davidben@google.com>