Re: [PoC] Let libpq reject unexpected authentication requests

v2 rebases over latest, removes the alternate spellings of "password",
and implements OR operations with a comma-separated list. For example:

- require_auth=cert means that the server must ask for, and the client
must provide, a client certificate.
- require_auth=password,md5 means that the server must ask for a
plaintext password or an MD5 hash.
- require_auth=scram-sha-256,gss means that one of SCRAM, Kerberos
authentication, or GSS transport encryption must be successfully
negotiated.
- require_auth=scram-sha-256,cert means that either a SCRAM handshake
must be completed, or the server must request a client certificate. It
has a potential pitfall in that it allows a partial SCRAM handshake,
as long as a certificate is requested and sent.

AND and NOT, discussed upthread, are not yet implemented. I tied
myself up in knots trying to make AND generic, so I think I"m going to
tackle NOT first, instead. The problem with AND is that it only makes
sense when one (and only one) of the options is a form of transport
authentication. (E.g. password+md5 never makes sense.) And although
cert+<something> and gss+<something> could be useful, the latter case
is already handled by gssencmode=require, and the gssencmode option is
more powerful since you can disable it (or set it to don't-care).

I'm not generally happy with how the "cert" option is working. With
the other methods, if you don't include a method in the list, then the
connection fails if the server tries to negotiate it. But if you don't
include the cert method in the list, we don't forbid the server from
asking for a cert, because the server always asks for a client
certificate via TLS whether it needs one or not. Behaving in the
intuitive way here would effectively break all use of TLS.

So I think Tom's recommendation that the cert method be handled by an
orthogonal option was a good one, and if that works then maybe we
don't need an AND syntax at all. Presumably I can just add an option
that parallels gssencmode, and then the current don't-care semantics
can be explicitly controlled. Skipping AND also means that I don't
have to create a syntax that can handle AND and NOT at the same time,
which I was dreading.

--Jacob