fix cldeadlock crashes

[markhannum]

Fixes three crashes surfaced by running the cldeadlock test in a loop against a 3-node cluster.

Dangling clnt->thd. A pooled sql worker leaves clnt->thd pointing at its per-thread state after finishing a query; once signal_clnt_as_done hands the clnt back to the event thread, the watchdog's long-running-client scan (reqlog_long_running_clnt) can read clnt->thd while the worker thread is torn down, and recover_deadlock_flags could dereference it from the post-done flush path. The worker now detaches itself before signaling done via a new clnt_detach_thd helper: it clears sqlthd->clnt under gbl_sql_lock and clnt->thd under clnt->sql_lk (the lock the watchdog scan holds), and skips the detach on a nested replay since the outer frame still owns the thd. The watchdog scan also snapshots clnt->thd once instead of re-reading it, recover_deadlock_flags uses the TLS sql_thread instead of clnt->thd, and osql_clean_sqlclntstate preserves in_replay_nested across the osql bzero since it is call-stack state, not transaction state.

SEGV in newsql_write_postponed_evbuffer. run_stmt set postponed_write unconditionally, so if saving the postponed row failed (or produced no row), newsql_send_postponed_row later dereferenced a NULL/stale pointer. run_stmt now only claims a postponed row when send_row(postpone=1) actually succeeded, and newsql_send_postponed_row NULL-checks before sending.

Abort sending startgen with rqid 0. After a master swing, a commit retry could run osql_send_commit_logic on a session that had already been ended: the clnt_check_bdb_lock_desired path in osql_sock_start_int called osql_end even when OSQL_START_KEEP_RQID was set, unregistering the session and zeroing osql->rqid, and osql_sock_commit then retried the commit anyway, sending OSQL_STARTGEN with sid 0 and tripping the sid==0 assertion in osqlcomm_rpl_type_put. The lock-desired path now skips osql_end when keep_rqid is set (matching the existing post-send failure path), and osql_sock_commit only retries the commit logic when the session is still registered (osql->rqid != 0), otherwise failing the transaction so the client retries.

[roborivers]

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

The first 10 failing tests are:
sc_truncate_multiddl_generated [db unavailable at finish] **quarantined**
sc_resume_logicalsc_generated **quarantined**
consumer_non_atomic_default_consumer_generated **quarantined**
sc_downgrade [timeout] **quarantined**

[roborivers]

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

The first 10 failing tests are:
sc_resume_logicalsc_generated **quarantined**
ssl_san
consumer_non_atomic_default_consumer_generated **quarantined**
ssl_set_cmd
ssl_prefer
ssl_dbname
sc_downgrade [timeout] **quarantined**
queuedb_rollover_noroll1_generated [timeout] **quarantined**