Add a schema delta to drop unstable private read receipts.

[clokep]

Follow-up to #13653, we need to somehow still handle unstable private read receipts that are in the database so that we don't leak them to other users.

As they're unstable it makes more sense to either:

1. Drop them completely.
2. Migrate them to the stable prefix.

I chose to just drop them since it is an unstable feature, but could be convinced to migrate them.

[clokep]

CC @SimonBrandner.

[erikjohnston]

Hmm, I don't think we have indices on receipt_type, so this might take a while?

[clokep]

The closest I can do to test this is a SELECT over the same rows, each query in that case takes a few seconds on matrix.org:

matrix=> EXPLAIN ANALYZE SELECT * FROM receipts_linearized WHERE receipt_type = 'org.matrix.msc2285.read.private';
                                                                   QUERY PLAN                                               
                    
------------------------------------------------------------------------------------------------------------------------------------------------
 Gather  (cost=1000.00..395762.79 rows=682557 width=166) (actual time=1.551..3131.541 rows=687402 loops=1)
   Workers Planned: 2
   Workers Launched: 2
   ->  Parallel Seq Scan on receipts_linearized  (cost=0.00..326507.09 rows=284399 width=166) (actual time=1.121..3027.560 r
ows=229134 loops=3)
         Filter: (receipt_type = 'org.matrix.msc2285.read.private'::text)
         Rows Removed by Filter: 3531513
 Planning time: 0.149 ms
 Execution time: 3149.429 ms
(8 rows)

matrix=> EXPLAIN ANALYZE SELECT * FROM receipts_graph WHERE receipt_type = 'org.matrix.msc2285.read.private';
                                                                QUERY PLAN                                                                 
-------------------------------------------------------------------------------------------------------------------------------------------
 Gather  (cost=1000.00..372414.02 rows=665573 width=131) (actual time=0.923..2521.685 rows=687399 loops=1)
   Workers Planned: 2
   Workers Launched: 2
   ->  Parallel Seq Scan on receipts_graph  (cost=0.00..304856.72 rows=277322 width=131) (actual time=1.034..2427.001 rows=229133 loops=3)
         Filter: (receipt_type = 'org.matrix.msc2285.read.private'::text)
         Rows Removed by Filter: 3531484
 Planning time: 0.764 ms
 Execution time: 2539.453 ms
(8 rows)

I think having a slightly slow start-up time for this is a reasonable compromise to avoiding a background update (and needing the additional filtering code while that runs).

This was the conclusion we came to for SimonBrandner#1 (comment) (and there was a team discussion I can dig up).

[SimonBrandner]

Thank you for another private read receipts catch - I really appreciate it!