Postgres source: slot startup check and feedback inaccuracies

[lukashes]

Grouped findings from the stability review, all in the replication path.

• The slot existence check (createSlotIfNotExists) queries pg_replication_slots by name only. A same-named slot in another database, or one created with a different plugin, passes the check, then START_REPLICATION fails on every start: a crash loop the supervisor cannot heal. Add AND database = current_database() and verify plugin = 'pgoutput' with a clear error.
• The drain loop swallows receive errors (source.zig: self.protocol.receiveMessage(0) catch break). A real connection failure is masked, the batch is returned as successful, and the error surfaces only on the next blocking receive. Log it at least.
• sendFeedback sends client_time as Unix seconds, but the protocol expects microseconds since 2000-01-01. pg_stat_replication.reply_time shows garbage. Cosmetic, still worth fixing while nobody depends on it.