Fix division by zero in arm paths with repeated configurations

[llm-learner]

Problem

Repeated joint configurations give adjacent path points the same cumulative length. Both interpolation sites in ArmConfigurationPath divide by that zero-length span. The final-segment fallback can also select a duplicate segment when Reflexxes slightly overshoots, producing non-finite joint targets.

This addresses the failure reported in #383. That earlier report already suggested guarding the division; this change covers both interpolation sites and fully stationary paths.

Change

• Use the segment start when its length is zero in both interpolators.
• Complete a non-empty stationary path without creating a Reflexxes trajectory, while applying and exposing its final joint target.
• Preserve path indexing and the existing formula for positive-length segments; no epsilon threshold or waypoint removal.

Validation

• Four new regression tests cover stationary and single-point paths, duplicate leading/interior/trailing segments, final-segment overshoot, and normal/tiny positive displacements. All pass with the patch. The duplicate-segment cases reproduce the original numeric errors with NumPy errors enabled.
• All 21 existing tests in tests.test_arms_and_configuration_paths and tests.test_grippers pass in CoppeliaSim 4.1.
• Real before/after replay: place_cups, TRAIN episode 0, variation 1, held object mug1. Each run independently resets and replays the same 14-action prefix. The restored state, requested release action and first native planned path match exactly. That 50-point path has total joint-space length about 1.91e-6 and repeated points, including a zero-length final segment.

Original	Patched
Seven NaN joint targets; release not reached	Finite targets; release completes
Still holding mug1	Grasp list empty

The diagnostic guard stops the original run before NaN targets reach the simulator. The video uses real front/wrist renders with the stopped side frozen and non-real-time playback labelled. No policy inference is involved. This demonstrates the numerical fix, not completion of the whole cup-placement task (reward remains 0).

中文说明

路径中重复的关节配置会产生零长度片段，原版两处插值可能除零；Reflexxes 轻微越过终点时，末段兜底也可能命中零长度片段。本修改保护两处插值，并直接完成非空静止路径。对比视频中，原版产生 7 个 NaN 关节目标，松爪未执行；补丁版目标有限、松爪完成。这里只验证数值故障修复，不代表整个放杯任务成功。

Before / after video

Bilingual comparison / 中英双语对比（19.2 s）

pyrep_zero_length_before_after_zh_en.mp4

Download MP4 / 下载视频

The video is attached above, with a backup on a separate evidence branch. This PR contains only the code fix and regression tests.