impossible to deal with correctly.

Pointed out by:		Dong Feng

Fix sess mempool names

stats counters as worker-local unless they actually are.

need it during vcl_hash{}

in as paramters to shave a few source lines

Retire a tiresome testcase to .disabled.

Constify some VRT prototypes.

I'm not quite there yet, but in the end we will have two workspaces:
req->ws for req.* and resp.* and busyobj->ws for bereq.* and beresp.*

Various argument changes as consequence of the sess->req moves.

a run of the mill mempool.

do not run in production!

Allocate the workspace as part of struct req, and move sp->ws,
sp->http and sp->http0 to struct req.

Allocate the req->http[0] on the workspace, so that changes to
nhttp does not force us to dump the mempool.

It's lifetime is now controlled by the timeout_linger parameter, so that
the waiters won't have to think about it at all.

request header:  There is no point in going through cnt_done to do that.

Collect all the code in cnt_wait()

Add a new -3 return code from HTC which says that all we've seen so
far is white space.  Use the RFC2616 "LWS" definition of white-space.

Rename "sess_timeout" parameter to "timeout_idle" and describe it more
precisely.  Make it a double, so more precise values can be specified.
The sessions is considered idle even if we receive white-space.

Rename "session_linger" parameter to "timeout_linger", make it a double
and give it units of seconds like other timeouts.

Add new parameter "timeout_req" which controls how long time we give
the client to send the request headers, from the first non-white char.
Set it to two seconds.  Experimentation/Experience requested.

"timeout_req" also gives us positive control of any kind of
"slowlaris" attack scenarios.

I'm trying to group parameters more sensibly, since we sort them
by name, calling them all timeout_*, pool_* etc, makes some sense I hope.

for a request.

t_open and t_idle are session timestamps.

Have the waiters pass in a T(now) timestamp

throw the session at a worker thread the moment it becomes readable.

worker pool.

Drop the mtx arg to MPL_New(), it wouldn't help.

Store the pool name locally.

Fix wrong type on some stats counters.

Copy gauge counters from reliable values.

Add stats-counter for dry pool allocations.

Add MPL_Destroy() which tickles the guard thread into self-destroying
the pool.  Only allowed if no live objects.

I have not made up my mind about yet.