With PR 5170, I added logic that requires a EVP_CTRL_GCM_SET_IV_FIXED command be
issued before a EVP_CTRL_GCM_IV_GEN command. This matches OpenSSL's behavior.
However, OpenSSL also clears the flag enabling EVP_CTRL_GCM_IV_GEN after
EVP_CTRL_GCM_SET_IV_FIXED if EVP_CipherInit is called with a NULL key.
Otherwise, the flag retains its value. We didn't mirror this logic, and that
caused problems in OpenSSH unit testing. This commit aligns our logic with
OpenSSL's and adds a regression test to test_evp_cipher_aes_gcm for this case.
fb3c611275