Complete with a bonus macOS bug. JFC does anyone QA this shit?

Once we start drifting this way, it's more natural to let the counters drift for
the entire duration, only becoming precise at the end. Hopefully this should
give us reasonable progress output while still accelerating performance.

This reverts commit 4315c307.

Switching to relaxed atomics actually resulted in a performance *degradation*.
This is an alternate attempt at performance optimisation where we remove the use
of atomics in the hot path of set_insert(). A side effect of this change is that
we lose accuracy in reported set occupancy. Both the progress output and the
decision on when to expand the seen set are now approximate. However, the final
summary of seen states is still precise.

Github: related to #117 "thread-local set size"

Github: related to #203 "murphi2c produces broken print code"

Github: closes #203 "murphi2c produces broken print code"

This is a breaking behaviour, but it will make resolving #203 simpler.

Github: related to #203 "murphi2c produces broken print code"

Not yet used, but we will need these in order to emit bounds in a couple of
places.

Github: related to #203 "murphi2c produces broken print code"

When emitting a print statement for a Murphi put statement of a string, the
string was incorrectly escaped. The value in a Put node in this case is already
naturally escaped, so needed no further escaping.

This seems to have been broken more or less forever, but apparently C compiler's
accepted this as a non-prototyped function. This was exposed as an error by a
recent transition to Apple clang version 12.0.0 which defaults to C99, requiring
stricter rules around pre-prototyping functions. We should probably set a C
standard explicitly during compilation in testing.

The aim here is eke out a little extra performance from set_insert(). Thanks to
a blog post from Travis Downs [0] for the inspiration here.

  [0]: https://travisdowns.github.io/blog/2020/07/06/concurrency-costs.html

To avoid a dependency on Ninja, as it is not installed by default in Travis and
installing it takes valuable time.

Github: closes #202 "remove section ordering restriction"

This opens the way to things like a constant or range limit that uses a
previously defined function. Though this is not yet supported.

Github: related to #202 "remove section ordering restriction"

Github: related to #202 "remove section ordering restriction"

Github: related to #202 "remove section ordering restriction"

Github: related to #202 "remove section ordering restriction"

Github: related to #202 "remove section ordering restriction"

Note that this has the following user-visible effects:
  * White space in C code produced by murphi2c has changed slightly; and
  * XML produced by murphi2xml no longer nests declarations, functions and rules
    under their own sections under <model>, but instead just directly under
    <model>.

Github: related to #202 "remove section ordering restriction"

The XML output is now also considered legal when declarations, functions and
rules appear directly as children of the 'model' element. That is, we no longer
require them to be nested under 'decls', 'functions', 'rules', respectively

Github: related to #202 "remove section ordering restriction"

Github: related to #202 "remove section ordering restriction"

After some upcoming changes to relax section ordering restrictions, this kind of
node reordering will have some undesirable side effects. In order to pre-empt
this, we take advantage of the fact that fields are only being reordered to get
more optimal offsets. Since the offset of state variables is stored within them
and not calculated dynamically, we can simply update their (explicit) offsets,
and leave their ordering as-is, neatly side stepping this upcoming issue.

Github: related to #202 "remove section ordering restriction"

Github: related to #202 "remove section ordering restriction"

Github: related to #202 "remove section ordering restriction"

While this is currently rejected, we may consider this legal in future.

Github: related to #202 "remove section ordering restriction"

The idea is to move to a less error prone visitor pattern. By requiring
visitation dispatch to be implemented on classes, we ensure a new AST node
cannot be introduced without handling its visitation support.

This is already available in tempfile.

It seems armel [0], armhf [1] and mipsel [2] now use this to back
gettimeofday().

  [0]: https://buildd.debian.org/status/fetch.php?pkg=rumur&arch=armel&ver=2020.07.28-1&stamp=1596373663&raw=0
  [1]: https://buildd.debian.org/status/fetch.php?pkg=rumur&arch=armhf&ver=2020.07.28-1&stamp=1596373870&raw=0
  [2]: https://buildd.debian.org/status/fetch.php?pkg=rumur&arch=mipsel&ver=2020.07.28-1&stamp=1596378655&raw=0