We can now use the benchmarks to dump data structures. This can be
done for both QP and PH, and is not limited by workspace or the
varnishtest log buffer.

VCL failure is invoked if:
- no entries were added to a set
- a set was not compiled
- .compile() is called in a VCL sub other than vcl_init
- a numeric index is out of range (larger than nmembers)
- the conditions for UNIQUE or EXACT fail
- associated data to be retrieved (string, backend etc) was not added

If .match() or .hasprefix() are called with a NULL subject, it is
logged using tag Notice, but is not an error (return value is false).
This is because it may or may not be intentional to attempt a match
against an unset header.

The .matched() method now may have a select argument, and works
similarly to other methods with the f(INT n, ENUM select) signature,
except that it returns false when the select condition fails, but
does not invoke VCL failure. This makes it possible to check if
UNIQUE or EXACT may be used, and avoid VCL failure if desired.

Fix rounding for h2strings_avg while we're here.

The hash inner loop iterates no more often than necessary.

We also set min and max lengths for strings for each secondary hash,
so that misses may be found more quickly.

- Add PH stats
- Stats are visible in varnishstat curses mode at verbosity level
  debug.
- Rename stats for better readability.
- Stats specific to PH and QP have names with prefixed hash_ and
  trie_, respectively.

Benchmarks of this idea can lead to improvements, by about 10 ns
per match operation, which makes a difference in throughput. But
we keep things simple for now.

The math was wrong, and changing the hash function to correctly
compute mod Mersenne prime just made it slower, and didn't seem
to lower collision rates.

Hash table sizes are just the next higher power of 2. As I interpret
Thorup (2020), this is still strongly universal hashing.

For large sets, workspace could be too small.

Universal hashing has a sounder theoretical basis; in particular, it
doesn't have the dubious minimum hash table size below which a
perfect hash may not be possible, and which was set by trial and error.

For nearly all test data, universal hashing performs at least as
well or better. Especially better for sets with longer strings,
since the subject string is cast as an array of uint32_t, so the
hash is computed in fewer operations.

The only exception I've noticed is /usr/share/dict/words, which now
appears to have more collisions than under the previous approach.
But it appears likely that this only becomes an issue for sets that
are much larger than are probable for VCL use cases (in the 100,000
range), and if all of the sets' elements are tested for matches
about equally often (whereas real-world usage patterns tend to
match a subset much more frequently).

The VMOD does this during .compile(), and QP_Insert() is no longer
called during .add(). The .compile() call is now required in all
cases, and it must be called before .create_stats().

This is because QP_Insert() was not correctly rotating the trie
when a set has overlapping prefixes, and a shorter prefix was
added before the longer one. With sorted order, shorter prefixes
are always added first, so rotation is unnecessary.

This reverts commit afded0f5.