radard: cluster lead acceleration

[GavinnnK]

Fixes #33934.

Lead acceleration was tracked independently for each radar track, so switching between tracks on the same lead could make aLeadK jump or restart from zero.

This averages acceleration and decay time across nearby mature tracks and initializes new tracks from that estimate. Position, velocity, and vision matching remain per-track.

Verification:

• added tests for track switching, the clustering threshold, and acceleration continuity on the next frame
• Toyota and Hyundai radard process replay passed
• RAV4 mean aLeadK jump on track switches: 0.0969 -> 0.0299 m/s²
• Corolla TSS2 mean aLeadK jump on track switches: 0.0723 -> 0.0147 m/s²

[github-actions]

Process replay diff report

Replays driving segments through this PR and compares the behavior to master.
Please review any changes carefully to ensure they are expected.

✅ 0 changed, 66 passed, 0 errors

[adeebshihadeh]

Looks like it was thrown in an LLM.

[GavinnnK]

The two failing checks are unrelated to this change. Both fail while importing raylib on the stock ubuntu-24.04 fork runner because libGLESv2.so.2 is missing. This is the known fork-runner issue tracked in #38435 and #38503. The radar tests pass, and process replay reports 0 changed, 66 passed.

[adeebshihadeh]

The tests don't validate this change. If you don't understand the PR, please don't put it up.