feat: complete GPU footprint labels and backend benchmarks

[joeblau]

PR #41 added the order-flow stack, but #40 still lacked GPU text and a real browser comparison. Footprint bid×ask labels now use a bounded glyph atlas with instanced quads; Canvas2D remains the fallback. Separate RGB/alpha blending preserves opacity when the GL canvas is composited.

Adds a Chrome benchmark for both backends at 1,000, 50,000 and 144,000 retained depth points. The largest workload spans 24 hours at one snapshot per minute, 50 levels per side. Recorded WebGL2 pan/zoom held ~60 animation FPS, with p95 CPU submission of 12.2/12.8 ms; Canvas2D reached ~24 FPS. These are rAF/CPU measurements on an M3 Max, not compositor FPS or a guarantee at finer history resolution. Raw results and reproduction commands are in packages/chart/PERFORMANCE.md.

Validation:

• npm run coverage: 1,855 tests pass; 100% line, branch and function coverage.
• All package entry points import in Node.
• Chrome pixel checks at 1×/2× verify glyph bounds, color, opacity, atlas reuse and no GL errors.
• Browser backend benchmark: 90 frames per workload, backend and retained-history assertions.

Closes #40. Closes #36.