Rename multi-segment functions to two-concept names; polish docstrings

Naming pass: rename functions whose third+ segment is redundant or
implementation-detail, sticking to the codebase's preferred
``noun_verb`` / ``verb_noun`` two-concept idiom. Renames are atomic
across definitions, callers, and tests.

  is_penned_position        →  is_penned
  modulate_speed_near_sheep →  modulate_speed
  mecanum_kinematics_step   →  mecanum_step
  policy_forward_mean       →  forward_mean

Two-concept patterns like ``velocity_to_wheels`` / ``detections_from_scan``
/ ``make_strombom_predictor`` are left alone — they're idiomatic
converters / factories that read as a single concept, and the longer
form aids grep-ability.

Docstring polish:
* ``herding/config.py`` header drops the "previously lived as a
  module-level literal" historical framing — we ship as a single
  thing, so the refactor anecdote no longer earns its keep. The
  usage examples now mention both ``HERDING_WEBOTS`` and
  ``HERDING_MEC_WEBOTS`` presets.

126 pytest cases still pass.

Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>

training/bc/pretrain.py

@@ -55,7 +55,7 @@ def build_model(net_arch_pi, net_arch_vf, log_std_init: float,
     return model, env
 
 
-def policy_forward_mean(policy, obs_batch):
+def forward_mean(policy, obs_batch):
     """Return the deterministic mean action for an obs batch.
 
     SB3's ActorCriticPolicy routes ``forward`` through a Distribution
@@ -177,7 +177,7 @@ def main():
             ob_batch = ob_batch.to(args.device)
             act_batch = act_batch.to(args.device)
             optimizer.zero_grad()
-            mean_action = policy_forward_mean(policy, ob_batch)
+            mean_action = forward_mean(policy, ob_batch)
             loss, mse_val, cos_val = combined_loss(mean_action, act_batch)
             loss.backward()
             optimizer.step()
@@ -196,7 +196,7 @@ def main():
             for ob_batch, act_batch in val_loader:
                 ob_batch = ob_batch.to(args.device)
                 act_batch = act_batch.to(args.device)
-                mean_action = policy_forward_mean(policy, ob_batch)
+                mean_action = forward_mean(policy, ob_batch)
                 bs = ob_batch.size(0)
                 val_total += nn.functional.mse_loss(
                     mean_action, act_batch, reduction="sum",