maskgen: --mintime/--mincomplexity act as maximums, and ?b/?h/?H/?1-?4 masks are costed as 1

[bandrel]

Two of these are regressions against iphelix/pack@622dd92, so they affect this
fork specifically.

--mintime and --mincomplexity are maximums

(self.mincomplexity is None or mask_complexity <= self.mincomplexity) and \
(self.mintime is None or mask_time <= self.mintime) and \

Upstream has >= for both. As written, asking for a minimum silently returns
exactly the masks you were trying to exclude. minoccurrence and minlength
kept their >=, so only these two are affected.

?b, ?h, ?H and ?1-?4 are costed as complexity 1

getcomplexity handles only l u d s a. Upstream 622dd92 handles
l u d s a b h H 1 2 3 4 (added in "Added the rest of the built-in charsets"
and "Added support for custom character sets"), and the --custom-charset*-len
options are gone from the CLI here entirely.

The consequence is quiet and expensive: a hex mask like ?h?h?h?h?h?h?h?h?h?h
falls to the else branch, which prints a warning but leaves count
unmultiplied. The mask is costed as 1, so it looks instant, sorts to the front
of an --optindex run, and never counts against --targettime. With this PR
that mask correctly costs 16^10 and reports ~18 minutes at 1e9 keys/sec.

Uncostable masks (a literal in a .hcmask, or ?1 with no declared length) are
now rejected with a message rather than admitted with an understated cost.

Masks repeated across input files are not aggregated

self.masks[mask] = dict() is last-write-wins, but every occurrence is added to
total_occurrence. Loading two files that share masks understates coverage —
easy to see with the same file twice, which reported Masks coverage: 50% (6/12) before and 100% (12/12) after.

Runtimes truncate to zero

mask_complexity // self.pps means anything under a second reports 0s. At the
default 1e9 keys/sec that is every mask below a billion candidates:
?d?d?d?d?d?d?d?d (1e8) and ?l?l?l?l?l?l (3.1e8) both showed 0:00:00. That
makes --mintime/--maxtime unable to distinguish them and lets
--targettime accumulate zeros without ever tripping. Complexity stays integral
and the division is now floating point, which also resolves the
# TODO: Something wrong here, complexity and time doesn't match with estimated from policygen — getmaskscoverage summed complexity and divided once while
loadmasks divided per mask, so the two paths genuinely disagreed.

Smaller items

• --targettime wrote the overshooting mask to the output file before breaking,
  so the emitted set always exceeded the budget by one mask.
• if options.mintime: and friends discard a legitimate 0; --pps 0 was
  accepted and then divided by. Now is not None throughout, and --pps 0 is
  rejected.
• statsgen's Advanced Masks section hardcoded if count*100//filter_counter > 0 while the other three sections gate on self.hiderare, so sub-1% advanced
  masks were always hidden from the display regardless of the flag. The output
  file was and remains complete.
• Empty or fully filtered input divided by zero in both tools.
• statsgen never closed its output file, so a killed run could leave a
  truncated .masks.

Testing

Regression tests for each item are in
bandrel/pack as tests/test_pack.py. Happy
to split this into per-bug PRs if you'd rather review them separately.

Related: #9 (rulegen), iphelix#31 (the subset affecting the
Python 2 original).