Add CircuitLens feature-circuit attribution - #131

Summary

• add model-agnostic CircuitLens artifacts and tensor decompositions for upstream transcoder features, frozen-pattern attention head/source-token contributions, and downstream output logits
• run attribution through public Target and HookSession activation/gradient capture APIs, including graph-retaining activation capture
• preserve existing gradients on model parameters and caller-owned leaf inputs, including their original .grad object identity
• fail closed when a configured hook site is reached more than once, preventing ambiguous forward/backward pairing for reused modules
• validate aligned attention-head selections and non-negative, in-range artifact positions and logit indices
• make local-Jacobian, frozen-attention, and frozen-nonlinearity conventions explicit and serialize them with contributor scores
• keep transformer-lens and circuit-tracer in an optional circuit-lens integration extra rather than TDHook core dependencies
• add analytical and contract-focused coverage for all attribution paths and failure modes

Validation

• uv run pre-commit run --all-files
• uv run pytest tests -q (682 passed, 1 skipped)
• uv run pytest tests/attribution/test_circuit_lens.py tests/test_session.py --cov=src --cov-report=term-missing -q (CircuitLens: 100% line coverage)
• uv build

Closes #125