We only want to return the connection early to the waiter when the
request is empty. Correct the read timeout calculation to reflect
that.

Thanks to Stackpath for helping to debug this issue.

Fixes: #2492

The documentation on directors needs an update too.

If workspace was exhausted, vmod_blob would fail yet still leave a
reservation which would likely trigger a WS_Reserve() assertion
failure in later code trying to reserve the workspace.

Fixes #2488

Thank you to @jarro2783 for the report and @dridi for the analysis.

beeaa19c had left 32 bytes on the client
workspace for the /baz case, which was sufficient to write resp.http.x-of
for 4-byte alignment, but not 8byte alignment:

* "false\0"		=  6b
* "x-of: false\0"	= 12b

Let's see if this fixes #2482

As dicussed during bugwash.

Ref #2489

partially reverts 9701bc56

It's not ready for primetime yet.

... as it was the case before 69d45413
and as documented.

The motivation is to remove the reservation from req->ws during
delivery, but actually line delivery memory should not come from the
request space - as originally designed:

- We avoid requiring an obscure surplus of workspace_client for delivery

  - which is also allocated for every subrequest though not required there

- We get predictable performance as the number of IO-vectors available
  is now only a function of workspace_thread or esi_iovs (see below)
  rather than the amount of memory which happens to be available on
  the request workspace.

As a sensible side effect, we now also fail with an internal 500 error
for workspace_session and workspace_thread overflows in addition to
the existing check on workspace_client for completeness.

For ESI requests, we run all of the client side processing, which uses
the thread workspace, with V1L set up. Thus, V1L now needs its control
structure together with a small amount of io vectors as an allocation
on the workspace.

Real world observation has shown that no more than five io vectors are
normally in use during ESI, yet still we make this number configurable
and have a default with some safety margin.

For non-ESI requests and headers, we use all of the thread_workspace
for io vectors, as before.

As V1L does not necessarily reserve workspace any more, functions have
been renamed to better reflect the purpose:

V1L_Reserve -> V1L_Open
V1L_FlushRelease -> V1L_Close

We still check that nothing is left on the thread workspace after
vcl methods in vcl_call_method(), but we do not require the thread
workspace to be totally empty any more.

Inspired by: #2488

This needs a separate job. Also fix previous commit.