The complexity analysis can be highly nuanced. Sets with many strings
and/or long strings can in fact require more time for matching due
to locality effects, which depends highly on usage patterns. Prefix
matches can be longer for large sets since the depth of branching
in the trie becomes larger.

But these are considerations at the level of micro-optimization.
It's the difference between two- or three-digit nanosecond mean
times for the matches.

So it suffices to claim that matching is fast and scales well.

Benchmarks have shown that (collision buckets)/(set size) ranges
from 10%-25%, so there's no need to tweak the hash table size.

Implmentation-dependent stats are now described in an external
document.

Fix some VSC long-form documentation while we're here.

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.