docs: add explanation for using KD∈[0,1] to control lateral target (no new KB)

[Jasmineyangyangyang]

Motivation

• Clarify how to retain only KJ/KD/KT while enabling left-to-right lateral bias via KD without introducing a new KB parameter.
• Provide a concrete mapping from KD to a lateral target l_ref and explain the resulting cost interpretation to guide tuning.

Description

• Add docs/18.md documenting that KD is repurposed to represent both the lateral target position and the bias weight while keeping the three-weight cost structure unchanged.
• Show the mapping l_ref = -MAX_ROAD_WIDTH + 2.0 * MAX_ROAD_WIDTH * KD and the replacement of the old offset term abs(tfp.l[-1]) with abs(tfp.l[-1] - l_ref) normalized as raw_bias_error / (2.0 * MAX_ROAD_WIDTH).
• Describe the weight normalization and final cost composition using KJ, KD, and KT with KT = 1.0 - KJ - KD after renormalizing KJ and KD when their sum exceeds 1.0.
• Include tuning guidance and notes on the semantic tradeoff that KD controls both target side and follow-through strength.

Testing

• No automated tests were run for this documentation-only change.

---

Codex Task