Tune particle-filter noise well above the data-generation noise

When implementing a particle filter for continuous localization from noisy lidar, set the particle observation std and motion noise LARGER than the noise used to generate the data (here ~8x sensor std, ~3x control noise). Tight noise makes the likelihood too peaked and collapses particles onto a wrong symmetric mode early. To tell an implementation bug from inherent multimodality: sweep particle COUNT — if error shrinks as particles grow, the filter is correct and the residual is global-localization ambiguity (the weighted-mean state estimate averages two modes). Also: a per-seed sweep at a single low particle count looks catastrophic and misleads tuning; test noise settings across seeds AND particle counts together.