Improve ball filter stability

[schmidma]

Summary

• Fix ball filter association and velocity covariance handling.
• Use a chi-square 95% threshold for maximum matching cost.
• Add debug-only camera-matrix cache and projected-ball diagnostics for timing issues.

Motivation

Ball estimates were unstable under head motion. This keeps the filter math consistent and exposes timestamp/cache evidence without changing cache APIs or behavior.

Implementation

• Include percept covariance in hypothesis/percept Mahalanobis matching.
• Keep moving/resting covariance updates in the correct state dimensions.
• Log camera-matrix nearest lookup time deltas, cache window, estimated capacity, head orientation, and projected percept positions at debug level only.

Verification

• cargo nextest run -p ball_filter
• cargo test --doc -p ball_filter
• cargo clippy -p ball_filter --all-targets -- -D warnings