Hand the replication-ack wait to a background thread (port of #1963, with measurements)

[emelialei88]

Port of #1963 (3126 commits behind main, hand-ported rather than rebased) onto current main, plus an A/B measurement of what it actually buys.

Why. Block processors commit locally and then block in trans_wait_for_seqnum_int until every replicant acks. With the default 8 writer threads, that caps in-flight commits at 8 no matter how many clients are connected.

What. A dedicated seqnum-wait thread takes the wait; the block processor returns to the queue immediately. The ack still happens, so replicants stay bounded — unlike rep-sync none, which drops it and drove two nodes permanently incoherent inside one 5-minute run at 64 clients. Off by default (async_dist_commit), togglable at runtime.

Rather than carry #1963's copy of bdb_wait_for_seqnum_from_all_int — which predated durable-majority accounting, gbl_2pc, non_durable_retry and set_durable_lsn — the shared logic is factored out of main into bdb_wait_for_seqnum_finish and bdb_wait_for_seqnum_mark_incoherent.

Measured. 8-node btda cluster, 300s runs, single-row autocommit INSERTs, one table per client, inserts/sec:

clients	control	waiter on	rep-sync none	ctrl spread	waiter gain
4	7,375	7,553	10,309	0.23%	+2.4%
8	11,656	12,077	14,111	1.02%	+3.6%
16	12,146	12,455	13,722	0.81%	+2.5%
32	12,040	12,453	13,653	1.31%	+3.4%

Scope confirmed reachable: 726,674 commits enqueued vs 726,658 rows, zero inline fallbacks.

Risk / why this is a draft. The gain is +2.4–3.6% and flat across client counts, so writer-thread starvation was not the binding constraint — the working hypothesis going in. Raising async_dist_commit_max_outstanding_trans 8→32→128 does not help either. rep-sync none shows 12–35% of real headroom, but this is not the lever that reaches it. Opening for the port and the numbers; I would not merge it on a 3% return for ~500 lines of concurrent code that has to be kept in step with bdb_wait_for_seqnum_from_all_int forever.

Three latent bugs found and fixed during the port, which would have hit the original too: thrman_register(THRTYPE_GENERIC) made begin_clean_exit() wait forever on a thread that never exits; the waiter took BDB_READLOCK with no bdb lock slot and killed the master with [FATAL] bdb lock not inited in this thread; and gbl_set_coherent_state_trace defaults to 1, whose per-node-per-poll trace is catastrophic for a polling waiter.

[roborivers]

Cbuild submission: Success ✓.
Regression testing: Success ✓.

The first 10 failing tests are:
reco-ddlk-sql **quarantined**
consumer_non_atomic_default_consumer_generated **quarantined**
tunables
sc_downgrade [timeout] **quarantined**