Improve error handling of HMAC computations and SCRAM

Hi all,

This is a follow-up of the work done in b69aba7 for cryptohashes, but
this time for HMAC.  The main issue here is related to SCRAM, where we
have a lot of code paths that have no idea about what kind of failure
is happening when an error happens, and this exists since v10 where
SCRAM has been introduced, for some of them, frontend and backend
included.  \password is one example.

The set of errors improved here would only trigger in scenarios that
are unlikely going to happen, like an OOM or an internal OpenSSL
error.  It would be possible to create a HMAC from a MD5, which would
cause an error when compiling with OpenSSL and FIPS enabled, but the
only callers of the pg_hmac_* routines involve SHA-256 in core through
SCRAM, so I don't see much a point in backpatching any of the things
proposed here.

The attached patch creates a new routine call pg_hmac_error() that one
can use to grab details about the error that happened, in the same
fashion as what has been done for cryptohashes.  The logic is not that
complicated, but note that the fallback HMAC implementation relies
itself on cryptohashes, so there are cases where we need to look at
the error from pg_cryptohash_error() and store it in the HMAC private
context.

Thoughts?
--
Michael