Map generation and positioning estimation parameters

[Emboiss13]

Description

This PR refines map generation behaviour to fix coverage, realism, and target-placement issues introduced in #35.

Problem summary:

• Antennas were placed randomly, not uniformly, which can create coverage gaps.
• Obstacles had no minimum separation, causing adjacent walls and unrealistic environments.
• Target locations were generated “per map” rather than grid-based, which reduces training data consistency.

Key changes

• antennaFactory (or relevant module) now uses evenly spaced layout rules instead of pure random placement
  • improves prediction coverage and avoids clustered/empty regions
• obstacles now enforce a minimum separation distance
  • prevents wall overlaps and near-touching walls
• map generation now decouples environment layout from target placement
  • generate map structure (antennas + obstacles) first
  • Establish a cell grid for each map
  • place target positions across every grid cell (full calibration grid)
• ensures ML data can be collected for every location in the grid, not only random target points
• updated tests and/or generator parameters so behaviour is deterministic and reproducible

Dependencies

• matplotlib (plotting)
• fastparquet (parquet columnar data storage format)
• pyarrow (parquet columnar data storage format)
• renovation (floorplan generation)

[Emboiss13]

Pending:

• Make floorplan and environment PNG optional (specified on terminal user input)
• Print execution runtime
• How can we optimise this?
• Make target generator.
• How am I storing the occupied vs free space area?
• How are we labelling the data, and does the whole outdoor vs indoor classification even make sense?