perf(light): seed the sky from the heightmap instead of every open cell

[TheMeinerLP]

From Project Status → Open → Smaller items:

A full sky propagation still queues every open cell of a column rather than seeding from a
heightmap. Measured at 79.3 → 55.1 µs in the rebuild, with byte identity verified over 240 worlds.

seedSky queued every position that sees the sky — up to 98 304 entries for a 24-section chunk, each
popped again and expanded six ways only to find that all six neighbours are already at the full
level.

Lighting a position and queueing it are two different things, and this is where the difference is
largest. A position whose four horizontal neighbours are open at the same height can raise nobody:
above and below it the column is at the full level or blocked, and so is every neighbour. It still
has to be lit — the output and the border exchange read the levels — but it has no business in the
queue.

The walk down each column already finds where the sky stops, so it now keeps that. The positions that
need queueing then fall out as a range rather than a per-position test: a neighbouring column is
dark below its own sky bottom, so a column queues from its own bottom up to the deepest bottom among
its four neighbours. Above that line every neighbour is open. On a fully open chunk nothing is
queued at all.

The heightmap is one int per column, so it does not grow with the height of the chunk and is
allocated once with the propagator.

The part worth reading: it took three attempts to build a test that could fail

The first check of this change was that all 212 tests passed. They also passed with the queued range
deliberately shortened by one position — the plainest way to get this wrong. So the suite proved
nothing about it, and the change had no coverage at all.

attempt	terrain	shortened range caught?
the existing suite	flat, fully open or fully closed	no — on flat terrain the range is empty and the whole derivation is a no-op
varied ground height per column	stone rising from y = 0	no — see below
varied ceiling height, stone	air above and below a solid roof	no — see below
varied ceiling height, slab	a roof that stops the sky and passes light sideways	yes, 2 of 6 red

The second and third failures have the same cause and it is not obvious. The position where a
column's sky stops is the blocker, and a blocker that occludes every face takes no light from the
side either. So the topmost position of the queued range has nowhere to give its light, and getting
that edge wrong is invisible against solid stone no matter how varied the terrain is.

A slab stops the sky from above and accepts light from its four sides. That is the only shape that
separates a correct range from one position short of it — and it is exactly the case
SectionOpacity stores occlusion per face for.

SkyHeightmapSeedTest (new, 6 tests) compares against the algorithm this commit replaces, written
out in the test rather than called, so the oracle shares no code with what it checks. Four
configurations use stone and two use slabs; only the slab ones catch the off-by-one, and the class
javadoc says so, because the next person to add a case here will otherwise reach for stone.

./gradlew build is green.

What I expect the measurement to say, so it can be wrong

Running the same pair of classes through the Benchmark workflow, three forks, main against this
branch, with falcoReadStates and minestomFull as untouched controls in the same job.

Two predictions, stated before the numbers arrive:

1. propagateSky improves at every height, and propagate — block light, which does not go
   through seedSky — does not move. That second half is a control with an expected answer of zero.
2. The 4-section regression from perf(light): three changes to the propagation loop from item 2 of Open #36 is probably not fixed by this. I attributed it to the
   flat column plus seedSky, and only the queueing goes away here; the walk down each column
   still reads the flat array. If the regression survives, that attribution was half right and the
   walk is the remaining half.

[github-actions]

Test results

  201 files    201 suites   4m 5s ⏱️
  683 tests   683 ✅ 0 💤 0 ❌
2 058 runs  2 058 ✅ 0 💤 0 ❌

Results for commit 153d31b.