bugfix: fixed a race condition in the new conn queuing mechanism.

This issue appeared in our EC2 test cluster, which compiles Nginx with
`-DNGX_LUA_USE_ASSERT` and `-DNGX_LUA_ABORT_AT_PANIC`.

The lua-resty-redis test:

    === TEST 1: github issue openresty#108: ngx.locaiton.capture + redis.set_keepalive

in t/bugs.t [1] would produce core dumps in the check leak testing mode.

The backtrace for these core dumps was:

    #0  0x00007fd417bee277 in raise () from /lib64/libc.so.6
    openresty#1  0x00007fd417bef968 in abort () from /lib64/libc.so.6
    openresty#2  0x00007fd417be7096 in __assert_fail_base () from /lib64/libc.so.6
    openresty#3  0x00007fd417be7142 in __assert_fail () from /lib64/libc.so.6
    openresty#4  0x000000000050d227 in ngx_http_lua_socket_tcp_resume_conn_op (spool=c/ngx_http_lua_socket_tcp.c:3963
    openresty#5  0x000000000050e51a in ngx_http_lua_socket_tcp_finalize (r=r@entry=0x5628) at ../../src/ngx_http_lua_socket_tcp.c:4195
    openresty#6  0x000000000050e570 in ngx_http_lua_socket_tcp_cleanup (data=0x7fd419p_lua_socket_tcp.c:3755
    openresty#7  0x0000000000463aa5 in ngx_http_free_request (r=r@entry=0xbfaec0, rc=http_request.c:3508
    ...

Which was caused by the following assertion in ngx_http_lua_socket_tcp.c
with `NGX_DEBUG`:

    #if (NGX_DEBUG)
        ngx_http_lua_assert(spool->connections >= 0);

    #else

Thanks to Mozilla's rr, a recorded session showed that
`spool->connections` was `-1`.

Unfortunately, reproducing this case does not seem possible, since the
failure is due to the request cleanup (`ngx_http_free_request`). Here is
an explanation:

    -- thread 1
    local sock = ngx.socket.tcp()
    sock:connect()
    sock:setkeepalive() -- pool created, connections: 1

        -- thread 2
        local sock = ngx.socket.tcp()
        sock:connect() -- from pool, connections: 1

    -- thread 1
    -- sock from thread 1 idle timeout, closes, and calls
    -- ngx_http_lua_socket_tcp_finalize, connections: 0

        -- thread 2
        sock:setkeepalive() -- connections: -1
        -- ngx_http_lua_socket_tcp_resume_conn_op gets called, assertion fails

In order to avoid this race condition, we must determine whether the
socket pool exists or not, not from the
`ngx_http_lua_socket_tcp_upstream` struct, but from the Lua Registry.
This way, when thread 2's socket enters the keepalive state, it will
respect the previous call to `ngx_http_lua_socket_free_pool` (which
unset the pool from the registry).

[1]: https://github.com/openresty/lua-resty-redis/blob/master/t/bugs.t

src/ngx_http_lua_socket_tcp.c

@@ -4132,6 +4132,7 @@ static void
 ngx_http_lua_socket_tcp_finalize(ngx_http_request_t *r,
     ngx_http_lua_socket_tcp_upstream_t *u)
 {
+    lua_State                      *L;
     ngx_connection_t               *c;
     ngx_http_lua_socket_pool_t     *spool;
 
@@ -4185,6 +4186,19 @@ ngx_http_lua_socket_tcp_finalize(ngx_http_request_t *r,
             return;
         }
 
+        L = spool->lua_vm;
+
+        lua_pushlightuserdata(L, ngx_http_lua_lightudata_mask(socket_pool_key));
+        lua_rawget(L, LUA_REGISTRYINDEX);
+        lua_pushstring(L, (char *) spool->key);
+        lua_rawget(L, -2);
+        spool = lua_touserdata(L, -1);
+        lua_pop(L, 1);
+
+        if (spool == NULL) {
+            return;
+        }
+
         spool->connections--;
 
         if (spool->connections == 0) {