sql: clear clnt->thd before handing the client back - fix for 'cldeadlock' test

[markhannum]

clnt->thd points at the sqlthdstate the query ran under, which is alloca'd on the pool thread's stack. Nothing cleared it, so it outlives the frame.

cldeadlock reaches it by orphaning handles mid-statement. When the client vanishes send_columns() fails; for a non-select sqlinterfaces.c:4042 sets query_rc = CDB2ERR_IO_ERROR, and done_cb_evbuffer() returns on that value before setting clnt->done = 1. The connection is left with done == 0 and a stale clnt->thd, and the once-a-second watchdog scan dereferences it after the pool thread exits.

Adds clnt_detach_thd(), called before signal_clnt_as_done() in sqlengine_work_appsock() and sqlengine_work_lua_thread(). Skipped on nested replay, where the outer frame still owns thd. Not done on the RC_INTERNAL_RETRY path: enqueue_sql_query() has already cleared clnt->thd and the new owner may have freed clnt.

Fallout: reqlog_long_running_clnt() reads clnt->thd once rather than three times; recover_deadlock_flags() takes the sql_thread from TLS as recover_deadlock_flags_int() already does; osql_clean_sqlclntstate() preserves in_replay_nested across its bzero.

Fixes issue discovered by cldeadlock.

[roborivers]

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

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

[roborivers]

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

The first 10 failing tests are:
consumer_non_atomic_default_consumer_generated **quarantined**
sc_downgrade [timeout] **quarantined**