Add batched traversal inference server (Option A) behind opt-in flag

Implements the central inference server pattern: a dedicated GPU
process owns advantage/strategy/league networks, batches policy
requests across traversal workers via shared-memory tensor pool, and
returns logits. Workers route forward calls through InferenceClient /
NetworkProxy when traversal.inference_backend == "server".

Default remains traversal.inference_backend: local. The server
backend regresses iter time ~3.8× on the inspected default config
(small-model dispatch + sync-blocking traversal capping realized
batch at ~num_workers=8 instead of the bs=64-256 needed to amortize
IPC overhead). Keeping the implementation behind the flag lets us
re-enable when (a) model size grows, (b) per-worker interleaved
traversal lands, or (c) eval becomes dominant — see
docs/performance.md "Option A Bench Result and Structural Ceiling"
for the full diagnosis.

Plumbing included:
- inference_buffers.py: shared-memory tensor pool with slot
  management.
- inference_client.py: per-worker client + NetworkProxy adapter for
  the existing traversal.pyx call sites.
- inference_server.py: spawn-context server process with
  batch-window aggregation, weight sync, shutdown sentinel.
- bench_inference_backend.py: A/B between local and server backends
  with eval/checkpoint disabled.
- test_inference_server.py: round-trip and integration tests.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

configs/deep_cfr/default_server.yaml

@@ -0,0 +1,108 @@
+run:
+  experiment_name: deep-cfr-default-server
+  seed: 79
+  max_iterations: 1000
+  max_minutes: null
+  device: cuda
+  use_amp: false
+
+rules:
+  n_colors: 5
+  n_ranks: 9
+  min_rank: 2
+  n_handshakes: 3
+  hand_size: 8
+  expedition_penalty: -20
+  bonus_threshold: 8
+  bonus_amount: 20
+
+encoding:
+  derived_playability: true
+  slot_aware_playability: true
+
+network:
+  hidden_size: 512
+  num_layers: 3
+  activation: relu
+
+traversal:
+  traversals_per_player: 280
+  max_depth: null
+  max_nodes_per_traversal: 1000
+  regret_matching_epsilon: 0.0001
+  outcome_sampling_epsilon: 0.2
+  outcome_sampling_value_clip: 500.0
+  outcome_unsampled_regret: zero
+  cutoff_value_mode: score_diff
+  cutoff_rollouts: 0
+  cutoff_rollout_policy: random
+  cutoff_rollout_max_steps: 300
+  opponent_policy: average_strategy
+  strategy_sample_interval: 1
+  store_strategy_on_traverser_nodes: true
+  store_strategy_on_opponent_nodes: false
+  num_workers: 8
+  worker_chunk_size: 8
+  progress_every_traversals: 10
+  endpoint_depth_bucket_width: 100
+  endpoint_depth_bucket_max: 1000
+  inference_backend: server
+
+regret_matching:
+  all_negative_fallback: argmax_tiebreak
+
+training_weighting:
+  mode: lcfr
+
+self_play:
+  snapshot_every: 1
+  max_snapshots: 0
+  anchor_probability: 0.0
+  current_weight: 1.0
+  recent_weight: 0.0
+  older_weight: 0.0
+  anchor_weight: 0.0
+  recent_window: 5
+
+optimization:
+  advantage_updates_per_iteration: 512
+  strategy_updates_per_iteration: 512
+  advantage_batch_size: 1024
+  strategy_batch_size: 1024
+  learning_rate: 1.0e-4
+  weight_decay: 0.0001
+  grad_clip: 1.0
+
+memory:
+  advantage_capacity: 2000000
+  strategy_capacity: 2000000
+
+checkpoint:
+  save_latest: true
+  save_every: 100
+  progress_interval_seconds: 20.0
+  exact_resume: false
+
+evaluation:
+  eval_every: 25
+  games: 100
+  opponents:
+    - random
+    - passive_discard
+    - safe_heuristic
+    - safe_heuristic_loose
+    - safe_heuristic_strict
+    - noisy_safe
+  max_steps: 10000
+  on_max_steps: score_diff
+  batch_size: 64
+  device: trainer
+  num_workers: 4
+
+inference_server:
+  device: cuda
+  num_slots: null
+  max_batch: 256
+  batch_window_us: 200
+  weight_sync_every: 1
+  use_amp: false