0.3.0

Changed

• Replaced the img2unicode dependency with a native dual-colour terminal
  renderer (pixel_map/terminal_renderers.py). The new engine rasterises
  glyph templates (geometric for half/quad/braille/block, brightness-ordered
  Bayer patterns for ASCII) and selects the best character per cell via a fully
  vectorised Lab-distance optimisation, yielding a 4x-60x speed-up over
  img2unicode while keeping the exact render_numpy(image) -> (characters, fg_colors, bg_colors)
  contract. All nine renderer names are preserved: block, all, ascii, space,
  half, quad, braille, braille-bw, ascii-bw.
• Dropped Python 3.9 support (EOL since October 2025; no numpy wheel exists for
  both 3.9 and 3.13). Minimum is now Python 3.10.
• The native renderer now has a fast path that skips the Pillow resize when the
  matplotlib canvas dimensions already match the subpixel grid (achievable with
  --dpi 8), and a row-averaged binreduce for the common 2:1 cell-aspect case.
• ASCII renderer now uses brightness-ordered templates with Bayer dithering
  instead of Pillow-font rasterization. Characters are assigned templates
  based on predefined brightness values, with denser templates for darker
  characters.
• Replaced CARTO DarkMatter and Positron basemap providers with ESRI
  World_Dark_Gray_Base and World_Light_Gray_Base respectively

Fixed

• Fixed a ValueError: cannot reshape array on high-DPI (Retina) displays
  where the matplotlib canvas buffer is physically larger than
  get_width_height() reports. The reshape now uses
  get_width_height(physical=True) and derives correct dimensions from the
  buffer pixel count.
• Replaced the removed canvas.tostring_rgb() with canvas.buffer_rgba() for
  matplotlib 3.10+ compatibility.

Removed

• img2unicode (and its transitive heavy dependencies scikit-image,
  scikit-learn and the Linux-only n2 ANN library) are no longer required.
  pillow is now a direct dependency.