If we fail a stream early from VCL via return (fail) after sending a
request body, the h2 sess will be stuck waiting forever.

Spotted by:	dridi

the tcp-pool, and implement the tcp-pool with it.

It should now be trivial to implement other pools,
for instance UDS using this.

Filenames subject to change still (s/tcp_pool/conn_pool/ ?)

This anticipates some features planned for 6.0 that are presently
not yet implemented:

- UDS requires VCL 4.1.
- VCL variables local.socket and local.endpoint

On reception of a SETTINGS frame with a new value for
INITIAL_WINDOW_SIZE, we also need to adjust the flow-control window for
any active streams. rfc7540, section 6.9.2

Ref: 63807dd2

Found while working on https://code.uplex.de/uplex-varnish/varnishevent

the section/subsection/subsubsection structure was broken.

Regarding the 304 and beresp.ttl / beresp.grace / beresp.keep I am
missing a fourth level, input welcome.

Feel free to improve further!

in ban_lurker_test_ban() we remove completed bans from the obans list
so the ban list tail might not be present in the obans list.

To ensure that we stop at the tail, we now find our obans iteration
termination ban walking the bans list backwards until we find an
active ban, the head or the first obans element. In the latter two
cases we are done.

If there are any bans left to mark completed, the last active ban must
be contained in the obans list so we should terminate correctly.

Fixes #2556

Closes #2598

Fixes #2597

Merges #2600 slightly modified