Previously, a FetchError in vbe_dir_getfd() was too general to be
useful. Now we log an error with the same identifier as the
corresponding statistic:

* unhealthy
* busy (max_connections reached)
* fail (no connection for real)

For these, we also log the backend name because we have not yet issued
a BackendOpen VSL.

A failed WS_Alloc is logged as "out of workspace".

We use an implicit rule to tell make how to turn vsc into c and
manually define the vsc files in Makefile.am. The implicit rule tells
automake to add .vsc to SUFFIXES, which in turn allows us to add the
actual vscs to varnishd_SOURCES.

To avoid triple configuration of the list of built VSC .c and .h
files, we generate these in configure. AC_SUBSTing files into _SOURCES
is forbidden by autoconf for a good reason: Adding sources to the
build should remain an explicit process to avoid accidentally
compiling in code just because someone had files with a magic name
lying around.

As the implicit .vsc -> .c rule does not allow for additional
dependencies, we add in the dependency to vsctool using the
auto-generated list of c files.

BUILT_SOURCES is used to ensure that, for default make targets,
headers generated from vscs are always present before the actual
varnishd sources are being build, of which some depend on vsc
headers. For subsequent builds, include dependency tracking takes
over.

The Makefile rules are deliberately simple to hopefully cover all make
flavours available on the planet. Notice that, for example, if we
limited ourselves to gmake, we could use much simpler substitution
rules to generate VSC_GEN_C and _H instead of the de-tour via
configure.

Portability is also the reason for renaming all vsc files.

recycle connections between protocols.

Should fix distcheck

someone(tm) might revisit this again later

Otherwise we might see garbled output files

StringIO does not exist any more in python3, yet requiring 2.7 would
not pave the path forward, so try to be compatible with both.

Works for me on Python 2.7.9 and Python 3.4

I would appreciate if someone more fluent in serpentinous programming
language reviewed and/or rewrote this.

These are versions of H2_Send_Get/H2_Send_Rel that can be called while
already holding the lock.

Maintain stream 0 window limits via h2->req0.

****heads in the gcc&llvm projects who cannot cooperate on
usability features ?

Fixes #2449

This commit enforces that a HEADERS frame without an END_HEADERS flag
must be followed by a CONTINUATION frame on the same stream.

Sending interleaved/out of sequence header or continuation frames will
now be treated as a PROTOCOL_ERROR, as per rfc7540.

This also fixes the panics seen in #2387.

Fixes: #2387

Remove duplication to some extent by moving common return keywords
to a separate paragraph.

Add fail and vcl(label) keywords

Fix some typos

Since we decoupled the client and backend side, restarts are completely
unrelated to backend requests.

Related to #2405

req->restarts only gets incremented in cnt_restart, so it will never
exceed max_restarts here.

Related to #2405

max_restarts is the number of allowed restarts

Related to #2405