Replace munkres with lapjv for track assignment

[snejus]

Description

Fixes #5207.

This PR replaces the munkres library with lap (Linear Assignment Problem solver) for computing optimal track assignments during the auto-tagging process. The main changes are:

• Remove munkres dependency and add lap and numpy dependencies
• Refactor the track assignment code to use lap.lapjv() instead of Munkres().compute()
• Simplify cost matrix construction using list comprehension
• Move config value reading outside of track_distance function to improve performance

The motivation for this change comes from benchmark comparisons showing that LAPJV (implemented in the lap library) significantly outperforms the Munkres/Hungarian algorithm for the linear assignment problem. See detailed benchmarks at: https://github.com/berhane/LAP-solvers

The change should provide better performance for track matching, especially with larger albums, while maintaining the same assignment results.

Testing Notes

• All existing tests pass without modification
• Track assignments produce identical results
• No behavioral changes in auto-tagging

[wisp3rwind]

This seems like the right call — with munkres being a pure-Python implementation, it's orders of magnitudes slower than the other libraries. Nice find that somebody already did the hard work of benchmarking all these packages!

The new dependencies look fine at a first glance: numpy is common (and I wouldn't be surprised if we depend on it already transitively). lap and munkres have a similar popularity on Github and have both seen updates about a month ago.

I left a small question about on code changes, otherwise this looks good to me; thanks!

[wisp3rwind]

Looks good to me 💯 (Assuming CI passes)