feat: compose component discovery with patching and best-layer steering

[alirezabayatmk]

Summary 📝

Makes two more flows composable. A discovery step can pick the components to patch and hand them to Patch/PathPatch at run time, and steering can apply only the best layer's direction, so both run as one pipeline instead of two with a manual hand-off.

Details

1. SelectComponents step and ComponentSelection artifact: rank a LogitAttribution result (by magnitude, sign, or most negative) and keep the top top_k or everything past a threshold. A modules= filter can keep heads only.
2. Patch / Ablate take targets_key= and PathPatch takes senders_key=, reading a ComponentSelection from Results at run time. The key is added to the step's reads, so a pipeline that fails to produce it fails pre-flight with a clear KeyError instead of silently patching nothing.
3. Intervene gets direction_layers ("all", "best", or a layer list): with direction_key it can apply only the best separating layer's direction instead of every recorded layer, which keeps deep models from degenerating.
4. Ablate captures target activations in layer order, so a ranked selection (or any out of order targets) patches correctly.

Checks

• Tested Changes: Verified on GPT-2, Pythia-1.4b, GPT-J-6B, Qwen2.5-7B, and Llama-2-13B.