Optimize CPU verifier and fix biome shift sampling

[sync0516]

Summary

• Fix the Monte Carlo prefilter's shifted-Z sampling order to match cubiomes: sampleDoublePerlin(shift, z, x, 0).
• Replace per-cell getBiomeAt() allocation with an exact staged mushroom-fields classifier: sample continentalness first, safely reject non-candidates, then run the remaining climate parameters and climateToBiome().
• Replace the full int state grid and worst-case preallocated flood queue with a reusable visited bitmap, compact candidate list, and dynamically growing queue.
• Add an opt-in CPU_ARCH=native setting for GCC/Clang CPU builds; the default remains portable.

Motivation

The production large-biome verifier covers 51,200 blocks. At scale 4 this is a 12,800 x 12,800 grid (163,840,000 cells). The current implementation allocates approximately:

• 625 MiB for the integer state grid
• 1,875 MiB for the flood-fill queue

That is about 2.44 GiB per CPU worker before allocator overhead.

The existing Monte Carlo callback also swaps the arguments used for the shifted Z coordinate. In a 262,144-point scale-4 comparison, this changed 15 prefilter decisions, including 11 where the old expression rejected a point accepted by the correctly ordered expression.

Benchmarks

Ryzen 5 7600X, Windows 11, GCC 16.2, large biomes, seed 664124959764, center (1600, 16000), range 51,200, minimum area 1,000,000, scale 4:

Measurement	Current	This PR
Peak private memory	2,505.6 MiB	82.3 MiB
Peak working set	662.2 MiB	58.1 MiB
Warm portable verification	8.93 s	8.26 s
Current portable vs this PR with CPU_ARCH=native	9.20 s	5.91 s

The returned center and area were identical: (2734, 6795), 72,710,512 blocks.

Correctness checks

• Exact staged classifier vs getBiomeAt(): 524,288 random points across normal/large biomes and scales 4/16, zero mismatches.
• Known large mushroom region: 1,661,598 continentalness candidates, zero exact-classifier mismatches.
• Eight full differential searches covering normal/large biomes, scales 4/16, hits, and misses: identical found state, center, and area.
• Default and CPU_ARCH=native CPU-only builds pass.
• -Wall -Wextra -Wpedantic produces no diagnostics.
• git diff --check passes.

A continentalness-only final classifier was tested and rejected: it produced 603 false-positive cells at scale 4 and 26 at scale 16 in the known region, which could inflate or connect islands.

Scope

This PR is intentionally CPU-only. It does not duplicate the GPU work already proposed in #28.