Add packaging metadata, CI, and unit tests for steering logic

[piyushbag]

Problem

The repository ships a minimal pyproject.toml alongside a legacy setup.py, and has no automated tests or CI:

• pyproject.toml declares only name and version. It lists no runtime dependencies, so pip install neural-steering produces an environment where import neuron_steer.core fails (no torch, transformers, or accelerate).
• setup.py duplicates name/version already present in the PEP 621 [project] table. On pip resolvers that fall back to the legacy setup.py develop path, pip install -e . fails outright.
• There is no test suite and no CI, so regressions in the core steering/selection logic are not caught.

Approach

Packaging (pyproject.toml, remove setup.py):

• Declare runtime dependencies (torch, transformers, accelerate), requires-python, description, readme, license, and project URLs so a plain install yields a working environment.
• Add an optional test extra (pytest) and a testpaths config.
• Remove the redundant setup.py and consolidate on pyproject.toml, which also fixes the legacy setup.py develop failure.

Tests (tests/test_steering.py, CPU-only, no model download):

• Steering multiplier logic in steer_neurons: multiplier=0.0 ablates only the circuit neurons (others untouched), 1.0 is a no-op baseline, >1.0 amplifies, and hooks are removed on context exit. Verified against a small stand-in module matching the model.layers[i].mlp.down_proj path that steer_neurons hooks.
• select_circuit: top-k, cumulative-threshold, and empty-input behavior.

CI (.github/workflows/ci.yml):

• Runs on push and PR across Python 3.9 and 3.11: installs CPU torch, editable-installs the package, runs an import smoke check on the public API, and runs pytest.

How this differs from the other open PRs

This PR intentionally touches a non-overlapping set of files:

• feat: offline safetensors ablation, async vLLM benchmarking, and polysemantic eval #11 — adds root scripts / data (apply_surgery.py, stress_test.py, canonical_indices.json, README).
• Fix ImportError: export public API from neuron_steer/__init__.py #12 — edits neuron_steer/__init__.py (public API re-export). This PR does not touch __init__.py; the tests import from neuron_steer.core so they pass with or without Fix ImportError: export public API from neuron_steer/__init__.py #12 merged, and are complementary to it.
• docs: drop references to API removed in the general-use cleanup #13 — README / examples/interactive_demo.py.
• fix: SyntaxError makes steering_comparison.py fail to parse #14, fix: CNA multiplier should be (1+alpha) so alpha=0 is baseline in steering comparison #15 — experiments/steering_comparison.py.

No file in this PR (pyproject.toml, setup.py, .github/workflows/ci.yml, tests/) is modified by any of #11–#15, so it can merge cleanly alongside them.

Test plan

pip install --index-url https://download.pytorch.org/whl/cpu "torch>=2.0"
pip install -e . --no-deps
pip install pytest
python -c "import neuron_steer.core as c; [getattr(c, n) for n in ('NeuronSteerer','Circuit','NeuronIdx','steer_neurons','select_circuit','compute_attribution')]"
pytest -q

Local run: 6 passed.