Speed up triangulation, especially hole bridging (~52% faster)

[mourner]

Closes #183. Makes earcut substantially faster on its primary workload — integer-coordinate polygons from Mapbox Vector Tiles — without any API or output-contract change. On a representative 132-tile fixture (119,680 polygons, 1.9M verts, 6.8% with holes), total triangulation time drops from ~1000 ms to ~480 ms (~52% faster).

Total-time win per zoom bracket

Bracket	Polys	Verts	main	PR	Win
z5–z8	60.2k	940k	~774 ms	~331 ms	~443 ms (−57%)
z9–z11	27.4k	595k	~200 ms	~126 ms	~73 ms (−37%)
z12+	27.3k	325k	~30 ms	~27 ms	~3 ms (−11%)

Want the z12+ number broken down further (e.g. per individual zoom), or is the three-bracket picture enough?

What changed

Hole bridging block index. findHoleBridge previously ran two unindexed O(n) scans of the merged ring per hole — 41.7% of total time. It now skip-walks an append-only block index (flat per-block bounding boxes, sized once and reused across calls, no per-call allocation):

• Leftward-ray scan skips any block whose bbox can't cross the ray (~97.5% of edge tests eliminated).
• Refinement scan skips any block that can't overlap the candidate triangle.

Combined, this took findHoleBridge from 41.7% → 11.8% of self-time and the holey-polygon bucket from ~948 ms → ~590 ms. Block size K=16 confirmed optimal by sweep.

intersectsPolygon diagonal prune. The split fallback's diagonal validity check now skips edges whose bbox can't overlap the (short) diagonal — bit-identical output, ~3% speedup.

z-order sort: linked-list merge sort → array quicksort. The cache-hostile pointer-chasing sortLinked is replaced with a cache-friendly array quicksort over node refs. Byte-identical output, removes code, ~2.5% speedup.

microoptimized earcutLinked: hoisted cheap reflex vertex check out of isEar / isEarHashed so that a common path doesn't call non-inlined functions. ~6% speedup.

microoptimized isEarHashed: inlined point in triangle checks in isEarHashed — the calls weren't inline due to being over inlining budget. ~2% speedup.

Faster hole merging & point filtering. filterPoints now heals only the dirty window around each bridge/diagonal cut (O(window)) instead of re-walking the whole ring per removal (was quadratic on holey polygons), with one upfront full-ring collapse after hole elimination to preserve global cleanup. ~4%

Dropped the "skip next vertex" sliver heuristic. It cost ~10% (+55% inner-loop iterations to relocate ears) for no quality benefit on real data — triangle shape is irrelevant to GPU fill rendering. ~5%. Adjusts a few canonical fixture outputs and loosens some tolerances on invalid/degenerate-prone polygons.

Minor cleanup. cureLocalIntersections skips its trailing full-ring filterPoints when it cured nothing (a guaranteed no-op on already-filtered rings). Removed steiner field from the Node shape since this is a rare path that shouldn't bring the cost to the common one.