[PATCH] Support pg_ident mapping for LDAP

Hello,

There was a brief discussion [1] back in February on allowing user
mapping for LDAP, in order to open up some more complex authorization
logic (and slightly reduce the need for LDAP-to-Postgres user
synchronization). Attached is an implementation of this that separates
the LDAP authentication and authorization identities, and lets the
client control the former with an `ldapuser` connection option or its
associated PGLDAPUSER envvar.

This isn't as useful as, say, authorization based on LDAP attributes or
group membership, but it seems like a necessary step towards that
feature, since we'll need to separate authn and authz anyway. This
provides some feature parity with other auth methods like gss, and it
solves the "let anyone who can authenticate against LDAP connect as X
role" use case trivially.

There is precedent for allowing the DBA to choose whether to map a bare
username or the "full" identity expansion -- see for example
include_realm=1 for gss and clientname=DN for cert -- so I added an
ldap_map_dn=1 option which can be used to map the whole DN. (I'm not
entirely convinced that it's useful, but maybe there are some
deployments that put authorization information into the LDAP topology,
like "everyone in this particular subtree is an admin".) Unlike
include_realm, this is only allowed with an explicit map. I don't
anticipate people using a full DN as a database username, and I'm
worried that doing that without normalization could cause some major
confusion and/or security problems.

When using a newer client with an older server, the new `ldapuser`
option will cause a connection failure. For the case where PGUSER and
PGLDAPUSER are identical, that failure is technically unnecessary, and
I briefly considered stripping the `ldapuser` option from the
connection string in that case so that we could have wider
compatibility. I now think that's a bad idea, because suddenly
introducing a hard connection failure (or new success) just because
your PGLDAPUSER variable changed would be a major support hazard. The
TODO is still in the code to remind me to have the conversation.

WDYT?

--Jacob

[1] https://www.postgresql.org/message-id/94f6b945f9ca8cabe2b9d2a38ec19dca6f90a083.camel%40vmware.com