Add a stroke rendering regression test over a type/head/tail matrix

[frankslin]

The testKage cases in test/index.js compare three hand-written glyphs
against inlined polygon coordinates. That is precise, but narrow: between them
they exercise 10 distinct (stroke type, head type, tail type) combinations, all
in mincho with kUseCurve off. Stroke types 3, 6 and 7 are never drawn, and
gothic is never rendered at all.

This PR adds test/strokes.js, which trades precision for breadth. It renders a
deterministic matrix and compares a fingerprint of each result against a
committed golden file:

• stroke types {1,2,3,4,6,7} x 9 head types x 15 tail types x 4 control point
  geometries x both shotai
• a kUseCurve sample over the curve-bearing stroke types (that code path in
  cdDrawCurveU is currently untested)
• whole-glyph cases covering buhin expansion, the two-segment stretch mode
  (sx > 100) and the 0:97 / 0:98 / 0:99 transforms

7,614 cases, about 0.3 s. npm test now runs test/index.js and then
test/strokes.js.

The fingerprint is (polygon count, point count, sha1 of the coordinates),
stored in test/strokes-golden.tsv. It only answers "did the output change",
never "is the output correct" — deciding the latter is still a visual job. When
output changes on purpose: UPDATE_GOLDEN=1 node test/strokes.js, then read the
diff. Each changed line is one stroke whose shape moved, so a diff that is
larger than expected is itself a signal.

Design notes

One geometry runs right-to-left and upwards. The README lists several
head/tail shapes that behave oddly when a stroke points that way, deliberately
left alone for now to avoid affecting non-kanji glyphs. Pinning them down means
that when they do get fixed, it shows up as an explicit golden-file update
rather than silently.

A case that throws records ERROR <message> as its expected value instead
of failing the run. Some combinations in the matrix are not meaningful for their
stroke type, and what matters is that the engine's answer to them stays put — a
combination that starts or stops throwing is exactly the kind of change this is
meant to surface. Non-finite coordinates are recorded the same way.

The corpus is entirely fixed literals — no RNG, no dates — so the golden
file is reproducible across machines and runs.

Baseline

0 throwing cases, 32 that draw nothing. The 32 are mincho straight lines with
head type 2 or 7 where the stroke is not horizontal, e.g. 1:2:2:50:20:50:80.
I checked those inputs against kamichikoichi/kage-engine and the behavior is
identical there, so they are recorded as-is rather than treated as a defect.

Sanity check that the test actually bites

Changing Mincho's kMinWidthY from 2 to 2.05 makes it report 2282 changed
cases; reverting restores a pass.

Notes

• No src/ changes, so no CHANGELOG entry — following the precedent of
  c9b8baa ("Add tests for rotation (0:99:2)"), which was also test-only.
• test/strokes.js passes eslint under the existing flat config.
• The golden file is 427 KB. If you would rather not carry that in the repo, I
  am happy to shrink it — truncating the hashes, dropping the rl geometry, or
  narrowing the tail-type list would each cut it substantially.

[kurgm]

Thank you for your contribution!