feat: RMA-style history-conditioned policy for sim2real adaptation

Added a temporal observation history buffer and 1D-CNN encoder so the
policy can implicitly infer environment parameters (mass, friction,
gear ratios, etc.) from recent (obs, action) dynamics.

Architecture:
  history window [(obs₀,a₀), ..., (obs_{H-1},a_{H-1})]
      → 1D-CNN HistoryEncoder → embedding (32-dim)
      → concat [current_obs, embedding] → MLP → action

Components:
- BaseRunner: history ring buffer, _push_history/_reset_history,
  augmented obs space (6 + H×7 = 76 with H=10)
- HistoryEncoder (src/models/mlp.py): 2-layer temporal Conv1d + GAP
- SharedMLP: optional history_length/raw_obs_dim/embedding_dim params;
  splits augmented obs, encodes history, feeds [obs, emb] to MLP
- TrainerConfig: history_length, embedding_dim fields
- All runner configs: history_length=10 by default
- Tests: encoder shape, model with/without history, config defaults

configs/runner/mjx.yaml

@@ -2,3 +2,4 @@ num_envs: 1024       # MJX shines with many parallel envs
 device: auto         # auto = cuda if available, else cpu
 dt: 0.002
 substeps: 10
+history_length: 10   # RMA-style: 10-step window of (obs, action) pairs

configs/runner/mujoco.yaml

@@ -2,6 +2,7 @@ num_envs: 64
 device: auto  # auto = cuda if available, else cpu
 dt: 0.002
 substeps: 10
+history_length: 10  # RMA-style: 10-step window of (obs, action) pairs
 
 # ── Sim2real: domain randomization ───────────────────────────────
 domain_rand:

configs/runner/mujoco_single.yaml

@@ -5,3 +5,4 @@ num_envs: 1
 device: cpu
 dt: 0.002
 substeps: 10
+history_length: 10

configs/runner/serial.yaml

@@ -8,3 +8,4 @@ port: /dev/cu.usbserial-0001
 baud: 115200
 dt: 0.02                    # control loop period (50 Hz, matches training)
 no_data_timeout: 2.0        # seconds of silence before declaring disconnect
+history_length: 10           # must match training runner