Adjusted readme, unified fm and fs calculation. Added stability mode.

[Conscht]

Summary

This PR introduces three major improvements:

1. Improved README and documentation
2. Unified false split (FS) and false merge (FM) calculations
3. Added stability mode to the evaluation loop

Motivation

• The README previously did not fully explain how to run the package as a module or how to use the Python API.
• FM/FS calculations were implemented via two separate code paths, leading to duplicated logic.
• Following the paper’s recommendations, models should be evaluated across three independent runs to assess robustness.

Documentation (README)

• Adds a clearer installation section, including a uv-based workflow.
• Expands CLI usage examples.
• Adds Python API examples using evaluate_file and evaluate_volume.
• Introduces a dedicated Stability & Robustness Mode section, documenting:
  • CLI flags: --stability_mode, --run_dirs
  • Expected output directory structure
  • The requirement of exactly three runs

Code Changes (evalinstseg/compute.py)

• Imports get_m2m_matches and introduces a shared helper (compute_m2m_stats) to compute FM/FS in a single place.
• Adds get_m2m_metrics(...) to compute (FM, FS, matches) in a unified way for many-to-many matching.
• Preserves compatibility with existing FM/FS entry points while reducing duplicated code.

Notes

• Stability mode is currently fixed to three runs, matching the evaluation protocol described in the paper.