Fix assorted problems in recovery tests

Fix assorted problems in recovery tests

In test 001_stream_rep we're using pg_stat_replication.write_location to
determine catch-up status, but we care about xlog having been applied
not just received, so change that to apply_location.

In test 003_recovery_targets, we query the database for a recovery
target specification and later for the xlog position supposedly
corresponding to that recovery specification.  If for whatever reason
more WAL is written between the two queries, the recovery specification
is earlier than the xlog position used by the query in the test harness,
so we wait forever, leading to test failures.  Deal with this by using a
single query to extract both items.  In 2a0f89cd717 we tried to deal
with it by giving them more tests to run, but in hindsight that was
obviously doomed to failure (no revert of that, though).

Per hamster buildfarm failures.

Author: Michaël Paquier