Align tiling scale math with MIN_MODEL_PX invariant and fix mismatched test intent

[Copilot]

This PR addresses all review-thread comments on PR #8 by tightening tiling scale rounding behavior, removing dead imports, and correcting a test scenario that did not exercise the path it claimed to validate.

• Tiling region size rounding

  • Updated plan_tiles() to compute region_w / region_h using floor-based rounding where required-scale math is applied, preventing crops from becoming larger than intended due to ceil(...).
  • This preserves the module invariant that a min_object_px target does not drop below MIN_MODEL_PX after resize/letterbox edge rounding.

• Test correction for overflow/truncation behavior

  • Fixed the test case in tests/test_tiling.py whose input contradicted its own intent (“very small min_object_px explodes the grid”).
  • Adjusted parameters so the test actually drives the high-tile-count / truncation path rather than collapsing to a full-frame tile.

• Code hygiene

  • Removed unused dataclasses.replace import from utils/tiling.py (including duplicate occurrence), eliminating lint noise.

# before (can overshoot region size)
region_w = math.ceil(model_w / required_scale)

# after (does not exceed required crop size)
region_w = math.floor(model_w / required_scale)