perf: eliminate O(n) hot-path scans in graph data management and tracker

[veillette]

GraphDataManager (issues 1 & 2):

• Replace dataPoints: Vector2[] with a fixed-size ring buffer (O(1)
  eviction vs. O(n) Array.shift reindexing on every oldest-point removal)
• Maintain running xMin/xMax/yMin/yMax updated incrementally in O(1);
  a full O(n) recomputeMinMax() rescan fires only when an evicted point
  was an axis extreme (rare once the buffer is settled)
• applyAxisRangesFromExtremes() replaces the previous updateAxisRanges()
  full-scan, keeping axis rescaling at O(1) on the 30 Hz tracking path
• addDataPoints() updates min/max inline per-point to avoid per-call
  function overhead in the tight loop

OpenCVTracker (issue 3):

• Split captureFrame() into drawVideoFrame() (GPU rasterise — fast) and
  readPixels(x,y,w,h) (CPU readback — expensive)
• track() now computes a search window around lastMatchCenter and calls
  getImageData only over that sub-region, reducing GPU→CPU pixel transfer
  by ~10–15× for typical template sizes at 30 Hz
• Falls back to a full-frame search when no prior match exists or the
  clamped window cannot contain the template
• lastMatchCenter seeded from initFromVideo() so the very first track()
  call benefits from the windowed path immediately

https://claude.ai/code/session_01E8HRtu2MvwhJYDGSu4ucPa