feat(m4): stage 2 SFT training script (LoRA r=16) + smoke (T4.1)

- src/tsmm/train/stage2.py: loads stage1 ckpt (optional), freeze_llm +
  enable_lora(r=16) on q/v_proj, masked LM loss on sft.jsonl, bs=4/grad_accum=8/
  ctx=1024/BF16/gradient checkpointing, ckpt every 2000 steps (encoder+projector
  state + peft LoRA adapter via save_pretrained).
- scripts/train_stage2.sh: wrapper with design defaults.
- Smoke verified: --max_steps 4, LoRA injected, finite loss, peak 3.29 GB
  (≤11GB budget), ckpt+adapter saved.

scripts/train_stage2.sh

@@ -0,0 +1,23 @@
+#!/usr/bin/env bash
+# Stage ② SFT training (T4.1). Loads stage ① ckpt, enables LoRA(r=16) on q/v_proj.
+set -euo pipefail
+
+DATA="${1:-data/sft.jsonl}"
+MAX_STEPS="${2:-10000}"
+STAGE1_CKPT="${3:-checkpoints/stage1/stage1_final.pt}"
+CKPT_DIR="${4:-checkpoints/stage2}"
+
+cd "$(dirname "$0")/.."
+.venv/bin/python -m tsmm.train.stage2 \
+  --data "$DATA" \
+  --stage1_ckpt "$STAGE1_CKPT" \
+  --ckpt_dir "$CKPT_DIR" \
+  --max_steps "$MAX_STEPS" \
+  --bs 4 \
+  --grad_accum 8 \
+  --ctx 1024 \
+  --lr 1e-4 \
+  --lora_r 16 \
+  --log_every 20 \
+  --ckpt_every 2000 \
+  --tensorboard