Deadlock re-mask still sends raw-graph hashes as bind parameters, which log_min_duration_statement writes into the store's own log #4035

Found while fixing #4012's review (finding 6), in the PR that follows up #4022.

What is left

That PR stops the deadlock re-mask from sending a row's raw TEXT back to the store as a bind parameter: the alert and finding writes now check the row by xmin, and the report write by the raw-hash test alone. Raw-graph HASHES still travel as bind parameters in three places in PgDeadlockRemask:

• ReportPageSql's cursor ($3, the last row's deadlock_hash);
• ReportUpdateSql's key ($2, the report's raw hash);
• AlertReportLookupSql's key array ($2, the raw hashes the alerts name).

With log_min_duration_statement set on the store, a slow statement's parameters land in the store's own log. A raw-graph hash is a test for the literals #4004 removed, so the store log would hold that test for a while after the rows themselves no longer do.

What a fix needs to decide

• The report walk needs a total order that does not include the hash (collection_time, server_id plus something stable on a compressed chunk).
• The alert lookup could be done in SQL from context_json itself, so the keys never leave the store.
• Whether it is worth it: the rows being rewritten still hold the same hash until the walk reaches them, and the pass finishes within hours on a measured 120k-row store.