refactor: migrate v0 prototype to Centaur 3-stage pipeline

Refactors the CRSM repository to move from v0 prototype scripts to the
modular "Centaur" training pipeline.

Changes:
- Archival: Move `train_full_crsm.py`, `hyperparameter_sweep.py`, and
  `rl_value_head_finetuning.py` to `scripts/archive/v0_prototype/`.
- Pipeline: Implement `stage_1_backbone.py`, `stage_2_dynamics.py`,
  and `stage_3_value_head.py` in `scripts/training/`.
- Configuration: Migrate all baseline configs in `configs/` from JSON
  to YAML and create main `training_config.yaml`.
- Evaluation: Add `benchmark_reasoning.py` and `verify_steering.py` to
  `scripts/eval/`.
- Notebooks: Add cloud-ready notebooks in `notebooks/cloud_training/`.
- Core: Update `crsm/cli.py` and `crsm/train.py` to support CLI arg
  parsing for config-based execution.

Author: Pomilon

configs/baseline_170m.json

@@ -0,0 +1,35 @@
+model:
+  vocab_size: 50257
+  d_model: 1024
+  d_state: 64
+  d_ffn: 4096
+  num_layers: 24
+  dropout: 0.1
+
+reasoning:
+  c_puct: 1.0
+  n_simulations: 50
+  temperature: 0.8
+  injection_rate: 0.05
+
+training:
+  batch_size: 8
+  seq_len: 1024
+  lr: 4e-4
+  backbone_epochs: 1
+  finetune_epochs: 1
+  finetune_lr: 1e-5
+  grad_accum: 4
+  use_amp: true
+
+dynamics:
+  dynamics_samples: 20000
+  dynamics_epochs: 5
+  dynamics_lr: 1e-3
+
+data:
+  data_dir: "data/text_corpus"
+  traces_path: "data/train_traces.jsonl"
+
+tokenizer: "gpt2"
+seed: 42

configs/baseline_170m.yaml

@@ -0,0 +1,35 @@
+model:
+  vocab_size: 50257
+  d_model: 512
+  d_state: 64
+  d_ffn: 2048
+  num_layers: 8
+  dropout: 0.1
+
+reasoning:
+  c_puct: 1.0
+  n_simulations: 25
+  temperature: 0.8
+  injection_rate: 0.05
+
+training:
+  batch_size: 32
+  seq_len: 512
+  lr: 6e-4
+  backbone_epochs: 2
+  finetune_epochs: 2
+  finetune_lr: 2e-5
+  grad_accum: 1
+  use_amp: true
+
+dynamics:
+  dynamics_samples: 10000
+  dynamics_epochs: 10
+  dynamics_lr: 1e-3
+
+data:
+  data_dir: "data/text_corpus"
+  traces_path: "data/train_traces.jsonl"
+
+tokenizer: "gpt2"
+seed: 42

configs/baseline_27m.json

@@ -0,0 +1,35 @@
+model:
+  vocab_size: 50257
+  d_model: 1536
+  d_state: 64
+  d_ffn: 6144
+  num_layers: 32
+  dropout: 0.1
+
+reasoning:
+  c_puct: 1.0
+  n_simulations: 100
+  temperature: 0.8
+  injection_rate: 0.05
+
+training:
+  batch_size: 4
+  seq_len: 2048
+  lr: 3e-4
+  backbone_epochs: 1
+  finetune_epochs: 1
+  finetune_lr: 5e-6
+  grad_accum: 8
+  use_amp: true
+
+dynamics:
+  dynamics_samples: 50000
+  dynamics_epochs: 3
+  dynamics_lr: 5e-4
+
+data:
+  data_dir: "data/text_corpus"
+  traces_path: "data/train_traces.jsonl"
+
+tokenizer: "gpt2"
+seed: 42

configs/baseline_27m.yaml

@@ -0,0 +1,43 @@
+model:
+  vocab_size: 1000
+  d_model: 128
+  d_state: 64
+  d_ffn: 512
+  num_layers: 2
+  dropout: 0.1
+
+reasoning:
+  c_puct: 1.0
+  n_simulations: 10
+  temperature: 1.0
+  injection_rate: 0.1
+
+training:
+  batch_size: 8
+  seq_len: 32
+  lr: 1e-3
+  backbone_epochs: 2
+  finetune_epochs: 2
+  finetune_lr: 1e-4
+  grad_accum: 1
+  use_amp: false
+  value_training:
+    rollouts_per_prompt: 4
+    epochs: 2
+    lr: 1e-4
+
+dynamics:
+  dynamics_samples: 100
+  dynamics_epochs: 2
+  dynamics_lr: 1e-3
+
+data:
+  data_dir: "data/text_corpus"
+  traces_path: "data/train_traces.jsonl"
+
+tokenizer: null
+seed: 42
+
+system:
+  device: "cpu"
+  seed: 42

configs/baseline_350m.json

@@ -0,0 +1,35 @@
+model:
+  vocab_size: 1000
+  d_model: 32
+  d_state: 16
+  d_ffn: 64
+  num_layers: 2
+  dropout: 0.0
+
+reasoning:
+  c_puct: 1.0
+  n_simulations: 2
+  temperature: 1.0
+  injection_rate: 0.1
+
+training:
+  batch_size: 2
+  seq_len: 16
+  lr: 1e-3
+  backbone_epochs: 1
+  finetune_epochs: 1
+  finetune_lr: 1e-4
+  grad_accum: 1
+  use_amp: false
+
+dynamics:
+  dynamics_samples: 10
+  dynamics_epochs: 1
+  dynamics_lr: 1e-3
+
+data:
+  data_dir: "data/text_corpus"
+  traces_path: "data/train_traces.jsonl"
+
+tokenizer: null
+seed: 42