feat(trajectory_planning): add BEV-only multi-sample trajectory scorer and shifted inference schedule

[FLagbusted]

PR title (Conventional Commits, per CONTRIBUTING.md semantic-pull-request check)

feat(trajectory_planning): add BEV-only multi-sample trajectory scorer and shifted inference schedule

---

PR body

Summary

Adds a Phase 1, training-free upgrade to the flow-matching driving policy,
scoped to exactly what was discussed in the 17/06 WG meeting: get
multi-sample scoring working against the BEV/map data we already have,
without building GoalFlow's full goal-point vocabulary + learned DAC
classifier yet (tracked separately as Phase 2).

Implements #75 .

What changed

• New: Model/model_components/trajectory_planning/trajectory_scorer.py
  — TrajectoryComplianceScorer, a zero-trainable-parameter wrapper that
  draws K samples from any BasePlanner and re-ranks them by drivable-area
  compliance (raw rasterized map pixel lookup) + kinematic comfort
  (acceleration/curvature bound violations).
• Modified: Model/model_components/trajectory_planning/flow_matching_planner.py
  — adds an optional timestep_schedule="shifted" inference mode
  implementing GoalFlow's (alpha·t)/(1+(alpha-1)·t) timestep warp.
  Default ("uniform") is byte-for-byte identical to current behaviour;
  this is purely additive.

Why

FlowMatchingPlanner (#40) generates one trajectory per forward call with
no scoring or compliance check. GoalFlow (arXiv:2503.05689, discussed this
week) shows multi-sample generation + re-ranking meaningfully improves
both safety compliance and final score (PDMS), but its full method needs a
goal-point vocabulary and a learned BEV semantic segmentation head neither
KITScenes nor L2D currently feed into this repo. This PR takes only the
part of that idea that needs nothing new: the rasterized map image already
produced by RasterizedMapEncoder's input pipeline (#55) is enough to do a
deterministic drivable-area check without training anything.

Code snippet — usage example

from Model.model_components.trajectory_planning import build_planner
from Model.model_components.trajectory_planning.trajectory_scorer import (
    TrajectoryComplianceScorer, ScorerConfig,
)

planner = build_planner("flow_matching", embed_dim=256, num_timesteps=64,
                        num_signals=2, egomotion_dim=256,
                        visual_history_dim=896)

scorer = TrajectoryComplianceScorer(
    planner, num_timesteps=64,
    config=ScorerConfig(selection="nearest", dac_weight=1.0, comfort_weight=0.5),
)

trajectory, ego_hidden, scores = scorer.sample_and_score(
    bev_features, visual_history, egomotion_history, map_input,
    num_samples=8, seed=0,
)

Testing done

• Smoke test (tests/test_trajectory_scorer.py) — unit-tested
  decode_trajectory_to_xy, project_xy_to_bev_pixel,
  drivable_area_compliance, kinematic_comfort_score, and
  TrajectoryComplianceScorer.sample_and_score against a fake planner
  satisfying the BasePlanner contract. Shapes, OOB-point handling, both
  selection modes, and the invalid-selection-mode error path all verified.
  Runs on CPU with no trained model or dataset required.
• Pending — hardware constraint: GPU inference test against real
  KITScenes data and a trained FlowMatchingPlanner checkpoint cannot be
  run for the next few days due to the contributor's machine being
  unavailable. Will update this PR with ADE/quality numbers once hardware
  is back — this is not blocking code or calibration review, only the
  final merge decision.
• Pending — calibration review: project_xy_to_bev_pixel not yet
  verified against the actual KITScenes/L2D renderer pixel convention.
  Requesting review from whoever owns the map renderer before merge.
• Pending — shifted schedule quality: timestep_schedule="shifted"
  quality impact on a real validation split not yet measured. Only the
  alpha=1.0 == uniform algebraic identity and basic schedule shapes
  were confirmed locally.

Checklist

• Conventional Commit PR title
• References the discussion/issue above
• Default behaviour of FlowMatchingPlanner.forward() unchanged
  (timestep_schedule="uniform" is the default)
• Smoke test passes (CPU, no GPU required)
• ruff check . passes
• Map-pixel calibration confirmed by renderer-owning reviewer
• ADE/quality validation on real KITScenes data — pending hardware