feat(map): add clean_keyframes.py to detect and prune non-discriminative keyframes

[xiaolefang-dm]

Summary

• Adds tool/clean_keyframes.py: a single tool that queries a built map with an independent
  probe bag, flags map keyframes whose top-K retrieval candidates repeatedly fail to spatially
  agree (perceptual aliasing rather than a genuine revisit), and prunes them into a new map copy.
• Also writes an HTML + images report explaining what was removed and why (per removed keyframe:
  its overall track record across the whole bag scan, plus worst-dispersion evidence queries).
• Detection method matches what was previously validated (topk=5, cluster_radius=1.0m,
  dispersion_threshold=0.4, min_participation=3); a keyframe is removed if it was ever the
  spatial outlier in a dispersed retrieval.

Test plan

• tool/clean_keyframes.py --map_path tinynav_db/maps/map_gt --bag_path tinynav_db/rosbags/bag_1970_01_01_08_05_15:
  486 queries, 16 dispersed, 28 keyframes removed (1120 -> 1092); report images + HTML verified correct
  (map_gt + the night probe bag are both available at
  https://huggingface.co/datasets/UniflexAI/rosbag_tinynav_vlad_eval)
• Draft: checking against additional maps before marking ready for review

Report example page

🤖 Generated with Claude Code

[github-actions]

cloc	github.com/AlDanial/cloc v 2.04 T=0.52 s (89.9 files/s, 262297.2 lines/s)

Language	files	blank	comment	code
JSON	3	8	0	129481
Python	15	742	357	4101
Markdown	4	155	6	426
C++	4	60	38	409
CSS	1	58	0	342
Bourne Shell	15	41	9	272
Dockerfile	1	31	32	204
HTML	1	6	0	107
make	1	20	15	62
TOML	1	4	0	59
CMake	1	4	0	17
--------	--------	--------	--------	--------
SUM:	47	1129	457	135480