[fmrico]

Closes #173.

• AmclRelocalizeMitigation makes up to attempts pose resets (default 5) at the pose from before the recovery, using RViz's covariance. Each attempt waits attempt_timeout (default 5 s) for convergence. The first attempt uses the exact pose; the others add an offset of up to max_offset (default 1 m), uniform over the disc. The robot is stopped every cycle.
• It does not reset if AMCL has already converged, and it checks quality_ratio like the evaluator. Without a robot pose it fails without publishing anything.
• rotation_speed and timeout are no longer used; setting them logs a warning.
• AMCLLocalizer: after an initial pose, reseeds use noise from the reset covariance × reset_refine_noise (0.25), decaying over reset_refine_time (3 s), so a stopped robot converges. Both parameters can be set to 0, which turns the refinement off; negative values are rejected.
• Tests: about 20 for the mitigation and 3 for the refinement. READMEs updated.

Kidnap test on the Kobuki playground (0.4, 0.6 and 0.9 m): all three recovered on the first attempt in 2–3 s, final error 0.03–0.09 m, and no re-trigger in the 15 s watched afterwards.