When WRW is used to send stuff to the backend 'req->t_resp' is not
an appropriate base for the timeouts.

We now have:
	VSL() -> send directly to global VSL buffer
	VSLb() -> send to buffered VSL
	VSLbt() -> send txt to buffered VSL

Fixes #1056

the "real streaming" code we are working on.

Fixes: #1092

the VSL coords which are now a property of struct http.

VSL logging.  More to come.

vbc's vsl_id.  When the vbc is returned, restore the original vsl_id.

Change WSLB() to VSLB() and have it take a busyobj as arg.

buffers wid with the vbc's vsl_id.  Flush the buffer when we
recycle/close the vbc.

Size controlled by param vsl_buffer which will replace shm_workspace.
The space comes out of request workspace.

locking:  Each pool fetches blocks of numbers to assign, then hands them
out one by one using the existing lock.

Session numbers are 32bit and start at zero on worker process startup.

Together with another 32bit number, local to the session, they will
be used to link VSL records into transactions.