I found myself needing this option to incrementally build a query and a
format on a large VSL compressed file. Because of the small nature of
the ncsa output I was working with and the narrow query I was refining,
and the decompression delay, it would take forever to complete this
pipeline:

    zcat '...' | varnishncsa -r - -g request -q '...' -F '...' | head -n 5

With the -k option it happens instantly. Alternatively, it could
probably have worked also with the -u option, but it wouldn't be a
one-liner change, and it would also result in a more complex command:

    zcat '...' | varnishncsa -r - -g request -q '...' -F '...' -k 5
    # vs
    zcat '...' | varnishncsa -r - -g request -q '...' -F '...' -u | head -n 5

The -k option would also generally be useful in "live" mode.

Better diff with the --word-diff --word-diff-regex=. options.

The vcc_feature parameter is in charge of translating param.set calls
from the alias name to the bit flag. On the other hand param.show will
defer to the generic alias display.

With this change we lose the "camel case" functions used to set those
bits through functions between mgt and libvcc. The functions were
renamed VCC_Opt_<lowercase> to look more appropriate than "all_lower"
or "ALL_UPPER".

It currently has no effect besides showing up in the CLI.

Refs #3269

Instead of simply resolving the original parameter and passing it on
pristine to its tweak, build an intermediary param spec that has all
the properties of the original with the name of the alias.

This enables special handling of deprecated aliases with their
replacements.

The debug stevedore only acts as a wrapper around the malloc storage, so
it proxies its initialization method. The allocation of av isn't tracked
so it leaks when it runs out of scope, along with the copies of values
assigned to the array.

Since a reference to aav is kept in stevedore::av, we can safely pass
direct references to arguments to the underlying malloc storage instead
of copies.

Spotted by asan.

We don't want to list ephemeral backends when we load a new VCL after
getting some s0 traffic.

While, before this change, subrequests would see the Transfer-Encoding
on the client side, I could not find a case where this caused any harm,
so this change, presumably, is only to prevent potential future issues.

Make sure subrequests to not see Content-Length / Transfer-Encoding

In other news: I've gotten much better at writing python.

Otherwise forks may attempt to run coveriy scan builds only to fail and
spam their owners.

Only momentarily until we understand and fix the newly introduced
issues.

This reverts commit f0ee94ec.

Ref #3813
Reopen #3761

There is no cancellation mechanism in the middle of an HTTP/1
transaction besides closing the session. If a backend closes
the connection before the end of the bereq delivery, it may
also send a 4XX or 5XX response that could be valuable for the
original client.

There are several things to take into consideration: it is
the responsibility of the backend implementation to raise the
bo->send_failed flag, if the backend replied something before
closing we shouldn't trigger a timeout, the connection must not
be recycled, and the object must not be cached.

Fixes #2332
Closes #3686