Unifdef SLAP_LIGHTWEIGHT_DISPATCHER
Use thread-specific data in pool_context()
Added native thread-specific data support
Concurrency tweaks:
  store conn->c_sd, don't use ber_sockbuf_ctrl to retrieve it.
  use per-thread free lists for operations, no mutexes needed.

Remove accidentally added #define LDAP_MAX_PENDING from last commit.

Remove ldap_int_thread_ctx_t.ltc_next.al (active list) left over from rev 1.75.

Tighten loop in pool_pause(): Omit unneeded ltp_pcond signals since rev 1.64.
Fix comment.

Not needed after ldap_int_thread_userctx_t.ltu_id was added and
_pool_wrapper took care not to update thread_keys[] during pauses.
Fix bogus thread_keys[].ctx comment.

- Move problematic mutex_unlock()-mutex_lock() into the unused #ifdef
  LDAP_PVT_THREAD_POOL_SEM_LOAD_CONTROL and add a FIXME comment inside.
- Delay ltp_starting--; until pool_wrapper().  Nonzero value could otherwise
  no longer be exposed when !defined LDAP_PVT_THREAD_POOL_SEM_LOAD_CONTROL.

Also plug a memory leak, set state vars even when unused for the sake of
debugging, and tweak some readability issues & data types.

Corrected rev 1.67 cvs comment: ltu_key[] is NULL-terminated _when not full_.

threads per pool, even when ltp_max_count <= 0 ("unlimited").
Keep ltp_max_count in range [0, LDAP_MAXTHR].

(setkey/getkey expected that, but purgekey could set a NULL in the middle.)
Added some checks for input key!=NULL.

API changes, need review - I'm not sure what's indented here:
- setkey(data=NULL, kfree!=NULL) searched as if intended to reset
  the key, but updated by setting the key.  Now always updates.
- setkey(key=<not found>, data=NULL) could return either success or
  failure.  Now succeeds iff (data == NULL && kfree == NULL).

"deleted item" mark.  Add DELETED_THREAD_CTX.
Also improve the hash function a bit, and make the hash unsigned.

ldap_pvt_thread_pool_purgekey() which may only be called during pauses.

Thus, also wait for pauses to finish before accessing thread_keys in
ldap_int_thread_pool_wrapper().  This may prevent pending tasks from
being started when a pause had been requested, which seems to have been
possible.  If that was a feature, we can split ltp_pause==1 in 2 states:
in pause (causes wait), and pause requested.

Also move 'thread_keys[].id = <thread id>' from pool_submit to
pool_wrapper.  Until pool_wrapper set the ctx as well, thread context
lookup would just return NULL anyway.

so a pause will work during states FINISHING and STOPPING.

Add missing waits and signals, and move waits in pool_wrapper().
Replace if(test) with while(test) when waiting for the multi-purpose
condition variable ltp_cond.

In ldap_pvt_thread_pool_submit(), when backing out of thread creation:
  ltp_pending_count '++' -> '--'.  Signal if there are no more threads.
In ldap_int_thread_pool_wrapper():
  if() -> assert() where false would result in eternal loop.