fix: stop the per-frame GPU readback that cost the wall 40x

[BernardJen]

The wall measured 1.4 fps on the split-flap board — about 700ms a frame — on an RTX PRO
4000 Blackwell with gpu_compositing enabled, where the same board runs at 118.9 fps in a
harness. #271's instrumentation is what made that visible:

saver                        mean    min    max   last   n  size
Split Flap (left)             1.4    1.2    2.0    1.2   4  2918x1200
Split Flap (right)            1.4    1.2    2.0    1.2   4  2918x1200

The cause, wrong in two compounding ways

Registered unconditionally, on the reasoning that "the observer is a no-op while
disabled". True of the observer, false of the readback — gl-base reads the frame before
calling anyone. So the wall paid full price with artnetEnabled: false the whole time, and
every stall fed a callback that did nothing.

Never rate-limited. The doc said observers were "expected to rate-limit itself" — which
they cannot: by the time one is called, the pixels have already been read. Art-Net sends at
1Hz while the readback ran at frame rate, so 59 of every 60 were thrown away.

Each notify is TILES_X * TILES_Y = 32 synchronous gl.readPixels. On a discrete GPU
each is a pipeline stall plus a PCIe transfer, and dual view runs two boards as two runtimes —
64 stalls per frame. 700ms ÷ 64 ≈ 11ms each, which is about right for that hardware.

Why I refuted this hypothesis yesterday

I tested exactly this and measured 4%. On Apple Silicon, where unified memory makes
readPixels close to a memcpy. The number was correct for this Mac and worthless as a
prediction for a discrete GPU on PCIe.

Measuring on the development machine was the mistake — not the reasoning. Worth recording,
because the same trap is waiting for the next GPU-cost question.

The fix

• the observer is registered only while Art-Net is enabled, re-evaluated on every settings
  save so toggling without a restart still works — the property the unconditional registration
  was protecting
• the readback is rate-limited to 1Hz inside gl-base, so the feature is usable on a
  discrete GPU even when it is enabled

Tests

9 new, 629 → 638. Both halves confirmed by breaking them:

Mutation	Tests failed
register unconditionally again	5
drop the readback rate limit	1

Two of these read gl-base's source, so they use projectRoot from the DOM helper —
import.meta.url is not a file URL under jsdom, which is a trap the helper exists to
document and which caught me here first.

Not yet confirmed

That this actually restores the wall's frame rate. The expectation is 1.4 fps → back toward
the harness figure, and #271's report is how that gets checked rather than assumed. The 1.4
fps baseline was captured deliberately before this change for exactly this comparison —
which is why the instrumentation shipped separately in the first place.

Filed #274 for the case this does not cover: Art-Net enabled on the wall has still never
been run, and the 1Hz limit being affordable there is arithmetic rather than a measurement.