Adds src/alpamayo_r1/metrics/smoothness_metrics.py with
compute_smoothness_metrics(pred_xyz, pred_rot, disable_summary,
planning_freq_hz) -> dict[str, Tensor] that reports per-batch RMS and
max magnitude of:

- jerk_lon  (longitudinal jerk, m/s^3)
- accel_lon (longitudinal acceleration, m/s^2)
- accel_lat (lateral acceleration, m/s^2)
- yaw_rate  (rad/s)
- yaw_accel (rad/s^2)

Why this is worth adding:

The kinematic derivations already exist in
finetune/rl/rewards/comfort_reward.py, but they are wrapped in a
within-bound boolean used as an RL reward. There is no easy way to
report the actual signal magnitudes during evaluation today, so a
researcher who wants to compare two checkpoints on "did this model
produce smoother trajectories?" has no metric for that.

This module surfaces the same physical signals as evaluation metrics
parallel to compute_ade / compute_fde. ADE/FDE answer "is the
trajectory accurate?"; smoothness metrics answer "is the trajectory
drivable?". Both are needed and they are independent.

Design notes:

- Pure additive change. comfort_reward.py is untouched. We deliberately
  duplicate the small _diff helpers so the metrics layer doesn't take a
  dependency on the RL training package; the docstring calls out the
  intentional duplication and explains why.
- Returns per-batch tensors of shape [B], matching the convention of
  every other metric in this module after summarize_metric. Per-key
  _std variants are added when N > 1 (consistent with compute_minade
  / compute_minfde).
- Yaw rate uses pi-wraparound handling so heading transitions across
  +/-pi do not report the spurious ~2*pi*freq_hz spike a naive diff
  would yield.
- Both peak and RMS are reported because they answer different
  questions: peak captures worst-case discomfort, RMS captures sustained
  effort; we have seen models that are good on one and bad on the other.

Also adds src/alpamayo_r1/metrics/test_smoothness_metrics.py with 9
pytest cases covering: returned keys + shapes, constant-velocity is
identically smooth, constant-acceleration recovers the input ax,
pi-wraparound is handled, _std gating on N and disable_summary, and
shape rejection on malformed inputs.

All 9 functional invariants verified locally against torch (no GPU /
HF needed) -- output:

    PASS: constant velocity gives v_lon=5, zero everywhere else
    PASS: constant accel ax=2 (interior mean = 2.000000)
    PASS: yaw wraparound (max |yaw_rate| = 0.400000, NOT ~62)
    PASS: emits 10 expected keys, shape [B]
    PASS: _std variants when N>1
    PASS: disable_summary drops _std
    PASS: constant-velocity reports near-zero on all signals
    PASS: rejects xyz wrong shape
    PASS: rejects rot wrong shape

Signed-off-by: lonexreb <[email protected]>