feat: add run-end terminal summary (#262)

[ny-shy]

What this PR does

When a training run finishes, the only way to figure out what happened is to dig through logs — how long it ran, what the final loss was, how many samples were processed, why it crashed. This PR adds a structured summary printed at the end of every run, regardless of whether it succeeded, was interrupted with Ctrl-C, or failed with an exception.

The summary includes:

• Outcome (success / interrupted / error)
• Wall-clock duration
• Final step and epoch
• Sample counts (processed / trained / skipped)
• Last recorded metrics
• Up to 5 error messages

On failure, the original traceback is preserved — the summary prints after the traceback, so neither is lost.

Two CLI flags control it: --summary/--no-summary (default: on) and --summary-json (default: off, emits JSON instead of text). With no flags, behavior is unchanged.

Files changed:

• New areno/cli/run_summary.py — data container + formatting + output, stdlib only
• New tests/test_run_summary_cpu.py — 21 CPU tests
• Modified trainer_config.py — added summary_enabled / summary_json fields
• Modified 4 trainers (policy_only / sft / dpo / ppo) — try/except/finally in fit(), data collection in training loop
• Modified train.py — CLI options, all 4 config construction sites updated
• Modified docs/cli/observability.rst — user documentation

Related issue

Fixes #262 Print a structured terminal summary when a run ends

Type of change

• New feature

How was it tested?

CPU unit tests on macOS (no GPU):

python3 -m pytest tests/test_run_summary_cpu.py -v
# 21 passed in 0.04s

Ran real SFT training on Kaggle T4 x2 to verify end-to-end behavior:

git clone -b feat/run-end-summary-262 https://github.com/ny-shy/AReno.git
cd AReno
pip install -e . --no-build-isolation

PYTORCH_ALLOC_CONF=expandable_segments:True areno train \
  --algo sft \
  --ckpt Qwen/Qwen3-0.6B \
  --dataset-path yahma/alpaca-cleaned \
  --dataset-loader-fn examples/sft/alpaca/dataset_loader.py \
  --model-hub hf \
  --tp-size 2 \
  --world-size 2 \
  --batch-size 1 \
  --mini-bs 1 \
  --max-prompt-tokens 128 \
  --max-new-tokens 64 \
  --max-steps 3 \
  --activation-checkpointing \
  --disable-thinking \
  --summary

Three scenarios were observed:

1. Success — 3 steps completed, outcome=success, metrics showed final loss=1.17, grad_norm=129.18, lr=9.9998e-07. Values matched train_stats in the training logs for the last step.

2. OOM — CUDA out of memory after 1 step (single T4, batch_size=2). outcome=error, error message captured and truncated in "Errors (bounded)" section, full traceback printed after the summary block. Summary correctly reported steps=1, samples=2 trained, and the metrics from step 0.

3. All data filtered — --max-prompt-tokens 16 --max-new-tokens 8 caused all 51,760 rows to be skipped. outcome=error, steps=0, metrics showed (no metrics recorded), samples all zero. Error message: ValueError: SFT dataset produced no valid training rows.

Summary output was correct in all three cases. No local GPU available, so full training was only validated on Kaggle T4 x2.

Checklist

• PR title summarizes the contribution
• Linked the related issue
• Existing tests pass (pytest tests/test_run_summary_cpu.py -v)
• New behavior is covered by tests
• Described test commands and hardware limitations
• CLI changes are additive and backward-compatible