v1.2.0: performance rewrite

- Batched descriptor writes (1-2 vkUpdateDescriptorSets per frame)
- Sorted best-fit pool allocator with adjacent-block coalescing
- Promotion queue + same-frame dedup in glyph cache
- Wave ballot compaction in task shader (no atomics)
- Precomputed dilation payload eliminates 2x sqrt in mesh shader
- BlobReader: reciprocal multiply + bitfieldExtract
- Debug-only Stats instrumentation (zero overhead in release)