docs(2d): how strong a blurred shadow gets, and how to test one

[sidorares]

Closes #287.

What I found

Shadows are not missing on the in-process JS X server. They paint exactly
what Xorg paints.

Drawing the issue's own snippet on both servers, with the same pinned font
so the glyphs are identical, gives byte-identical pixels — fillText and
TextLayout.draw alike, and the same for rects, paths, strokes and images
across blur sizes. Nothing in lib/ needed changing, so nothing in lib/
changed.

What the report actually found is how strong a blurred shadow gets. A
blur is coverage convolved with a gaussian, so a shadow reaches
shadowColor only where the shape casting it is wide compared with the
blur:

what casts it	shadowBlur	peak alpha
a 60x40 rect	8, sigma 4	1.00
a 60x40 rect	30, sigma 15	0.78
48px glyph stems	14, sigma 7	0.37

A five-pixel stem against sigma 7 keeps about 0.28 of its coverage,
which is the gaussian's own answer, and it is what a browser draws too. So
countPixels(ctx, band, '#ff0000', 90) returns 0 on a canvas whose red
glyph shadow is plainly visible — on any server, XQuartz included. The
comparison in the issue was a visual check on one server against an
exact-colour count on the other.

The picture

Same code, same font, two servers. Left is the in-process JS server with no
$DISPLAY; right is XQuartz. Differences outside the small captions: 21
bytes in the glyph-shadow band and 174 in the rect band, none larger than 1.
The captions differ because small-glyph rasterization is deliberately not
pixel-exact, which docs/xserver.md already said.

What changed

• docs/context-2d.md — a section on the peak alpha a shadow reaches,
  the three measured numbers, and what to assert instead of a colour: the
  shadow's own alpha on a transparent target, a difference between two
  places, or the gaussian's CDF profile.
• docs/xserver.md — server-side filtering is exact on the JS server
  even though rasterization is not, so a shadow that seems to vanish under
  a headless harness is an assertion problem rather than a missing request.
• test/shadow.test.js — hermetic: the wide-shape case pinned at 200/255
  and the 48px glyph case at about 94/255, with the reason in the comment.
• test/smoke-canvas.test.js — the same 0.784 asserted against Xorg.
  That cross-server equality is the claim the issue needed and nothing was
  checking.

For the react-x11 side

The font explorer's shadow is testable as it stands. Assert the shadow's
alpha rather than its colour: draw with fillStyle transparent so only the
shadow paints and read the alpha channel, or compare a pixel in the shadow
against one outside it. An exact #ff0000 count will keep returning 0 on
XQuartz too.