This parameter should ease the introduction of features that may later
be removed without being considered a breaking change. For features that
get promoted to either being always present or behind the feature param,
it sends a clear signal regarding when we consider a feature ready for
production. This could have been useful for HTTP/2 support for example.

The first experimental feature is the ability to drop thread pools,
which was behind the debug param.

Having EXPERIMENTAL in the code takes a lot of real estate, but exp is
already established as short for expiry.

This changes the default behavior of includes, matching the ESI language
specification. The onerror attribute seems to only accept one value, so
in its absence a failed ESI fragment delivery aborts the top request
delivery.

This new behavior requires the esi_include_onerror feature flag to be
raised to take effect.

In addition, it breaks the ESI object attribute format for persisted
caches.

Create a general list for holding "args to be dealt with later"
and use it for pretty much everything in that class.

Side-effect:  Multiple -T, -M and -P options are supported now.

Suggested by @gquintard.

We see linking failures sometimes, so hopefully that solves those.

Closes #3702

We grab the image from Red Hat's quay.io image repository, which is done
with the addition of an optional prefix parameter for distcheck jobs and
a hardcoded hack for packaging jobs.

To make this work, the aarch64 packaging jobs had to be generalized to
less specific "arm64" images.

But only fedora:latest has a packaging job.

Better diff with the --ignore-all-space option.

Presumably, to run images with fairly recent software?

Not all RPM-based distributions are EPEL consumers.

Better diff with the --ignore-all-space option.

Instead we can run packaging jobs directly from the target system's base
image.

The dist job is here to create a dist archive that serves as input for
packaging jobs. The distcheck jobs work directly from a git clone.

The fixed chunk size of one prevents testing the vxid cache behaviour.

in other words: Allow comments in the replay string

To match (pun intended) the naming convention in other places, keep
"match" for regular expressions matching.

Setting breakpoints in a debugger on select vsl events is tidious,
so add a simple facility to trigger abort() with a logexpect command.

Otherwise sanitizers suppressions may scroll the first expectation out.

Note to self: `process pNAME -winsz` is not documented.

TOSTRAND() creates a pointer with temporary lifetime which we must not
access outside of the expression where it is created.

Incidentally, this is what I had before a polish with good intentions,
which, as we all know, are not enough.

Noticed by Geoff, thank you

Ref ISO/IEC 9899:2018 §6.2.4 (8)
Ref 0c96fc65

Spotted by Coverity Scan.

Spotted by Coverity Scan.

Spotted by Coverity Scan: if we have a reservation, it must be the last
workspace allocation, so there must be one.