fix(patchworkpp): clear ringwise_flatness every ring (#69)

[LimHyungTae]

What

Move ringwise_flatness.clear() outside the if (!candidates.empty()) block in cpp/patchworkpp/src/patchworkpp.cpp so it runs unconditionally at the end of every ring iteration.

       if (!candidates.empty()) {
         if (params_.enable_TGR) {
           temporal_ground_revert(ringwise_flatness, candidates, concentric_idx);
         } else {
           for (auto candidate : candidates) {
             addCloud(cloud_nonground_, candidate.regionwise_ground);
           }
         }
         candidates.clear();
-        ringwise_flatness.clear();
       }
+      ringwise_flatness.clear();

Why

Reported by @KennethBlomqvist in #69. The previous placement only cleared ringwise_flatness when the current ring produced at least one revert candidate. If a ring finished with candidates empty, the accumulated flatnesses leaked into the next ring's temporal_ground_revert call and polluted its mean ± stdev statistics. Since flatness statistics differ across CZM rings (closer rings are flatter by construction), this biased the revert decision threshold (mu_flatness = mean + 1.5 * stdev, line 435).

Impact in practice is small because most rings DO produce candidates, so TGR fires every ring on KITTI. But the bug is real and the fix is one line.

Numerical impact — KITTI 00–10, full 23,201 frames, Patchwork++ paper protocol

Build	P	R	F1
v1.3.0 baseline	95.5494	97.1649	96.2886
This PR	95.5496	97.1710	96.2918

ΔF1 = +0.003 (within run-to-run noise). Confirms no regression.

Test plan

• One-line code change verified by reading the immediately-surrounding loop (the only callsite of temporal_ground_revert).
• pip install -v ./python/ from a clean conda env succeeds.
• Smoke test seq 00 first 50 frames → numbers within ±0.1 of v1.3.0.
• Full KITTI 00–10 sweep with --method patchworkpp --eval_protocol patchworkpp → table above.

Closes #69.