CI: 030_Classification_Optimization kills the CI kernel and gates the nmisp_py auto-sync

## Symptom
The `test ipynb on Google Colab` job (`conda_env_test.yml` → `test_ipynb_colab`) fails while executing `15_optimization/030_Classification_Optimization.ipynb`:

```
nbclient.exceptions.DeadKernelError: Kernel died
E0000 ... cuda_platform.cc:52] failed call to cuInit: INTERNAL: CUDA error: Failed call to cuInit: UNKNOWN ERROR (303)
```

## Root cause
- The `cuInit` line is a **benign warning** — there is no GPU on the GitHub runner, so TensorFlow/Keras logs it and falls back to CPU. It is *not* the cause.
- `030` **trains a TensorFlow/Keras model (cell ~101) and a PyTorch model (cells ~105–108)** in the same kernel. Loading both DL frameworks plus training on a CPU-only ~7 GB runner exhausts memory / trips a CPU-only crash → the kernel dies.
- This is a CI-environment limitation, not a logic bug; the notebook runs fine on Colab (more RAM, a GPU).

## Impact (the important part)
`update_nmisp_py` declares `needs: [test_ipynb_linux, test_ipynb_native, test_ipynb_colab]`. Because `test_ipynb_colab` stays red on `030`, **the `nmisp_py` auto-sync never runs**, so the Colab helper package (`kwlee2025cpp/nmisp_py`) silently goes stale — which is what broke the slider notebooks in Colab until a manual push.

## Options
1. **Skip `030` (and likely `035`/`036` Keras-MNIST) in CI** — via `TEST_IPYNB_IGNORE_FOLDER` or a pytest skip/marker. Simplest; accepts that DL-training notebooks aren't CI-testable on free runners.
2. **CI-guard the heavy cells** — wrap training in `if not os.getenv('CI'):` or drop to `epochs=1` / tiny data under CI, keeping some smoke coverage.
3. **Decouple `update_nmisp_py` from `test_ipynb_colab`** so a heavy-notebook failure can't silently staleness the Colab package.

## Related
A proposed static test (a notebook importing an `nmisp_py` helper module must include the Google Colab clone cell) would run in this same job — it passes today, but stays masked while the job is red on `030`.

🤖 Filed via Claude Code while debugging the Colab helper-sync gap.