- 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.

The new algorithm improves efficiency with iteration in place of
recursion, and in a number of other ways:

- Avoid searches into dead-end branches. The traversal of all branches
was done because of the overlapping prefix case -- "foo" and "foobar"
both in the set. Now we just search the tree for a match, but before
descending into the next branch, check if there are other branches at
which the current prefix matches a terminating node.

- Only do string comparisons when we hit a terminating node.

- Mark terminating nodes with a flag in the tree, so that we don't go
looking for the null byte in the strings table during the search.

While we're here, rename the flag for the nibble search as hinib --
non-zero if and only if we inspect the most significant nibble at that
node. Also remove some dead code from QP_Insert().

May be advantageous for loop/branch prediction.

Theoretically, this reduces the probability of collisions. Benchmarks
don't show much of a difference.

This adds the .compile() method to set objects, required for the
use of .match().

Docs for the .compile() method are currently incomplete.

strlen() is also cheap if it has a SIMD implementation, so we can
afford this optimization to reject some strings immediately.

Cannot be used for prefix matches.

For "quadbit patricia tries", inspired by the work of Tony Finch:
https://dotat.at/prog/qp/README.html

Radix 16 tries, examining a nibble at a time, to make the tries
smaller and reduce pointer chasing.

Only call strcmp() once, when a node is reached that must be either a
hit or a miss.