Refactor evaluation to use Inspect, bump package versions, rename package to steerability

[emiehling]

• Replace the Benchmark, metrics, and use-case evaluation modules with an Inspect AI evaluation package (steerability/evaluation/)
• Add as_inspect_model and ProviderOptions, a generation-only Inspect model provider over a steered pipeline, with a collator that batches concurrent requests (evaluation/batching.py)
• Add runtime_kwargs_solver for per-sample runtime kwargs, sample_scorer_from_inspect for using Inspect scorers as per-row rewards, and InspectSuite for defining task sets
• Add the SteeringEval runner (configurations x trials x suites), which resumes from the .eval logs and exposes results, runs_frame, samples_frame, and summarize_runs; rework viz_utils into evaluation/plotting.py over the summarized runs frame
• Move identity.py into core and add core/sweeps.py (expand_configurations, preflight, PipelineFactory) and core/scoring.py (SampleScorer); the sweep layer has no Inspect dependency
• Port to transformers v5 and bump the ML dependencies (trl 1.0, peft 0.20, datasets 4.7, accelerate 1.4, vllm 0.26 with trl[vllm], mergekit from a pinned commit); require Python 3.12
• Replace the plots extra with inspect and viz, declare merging as a uv conflict with inspect, all, dev, docs, and vllm, and commit uv.lock
• Rename the package, distribution, identifiers, and documentation from aisteer360 to steerability
• Add the .spipe serialization format (steerability/spipe/), which stores the recipe and, once steered, the frozen resolution in a content-addressed artifact store with a lock section; add SteeringPipeline.to_spipe() and SPipe.load, save, pipeline, verify, and thaw
• Add the freeze methods on controls (steer_fits, export_state, frozen_form, fit_identity) and the load_checkpoint and load_lora structural controls, which are the frozen forms of trained structural controls
• Add core/internals/model_layout.py (resolve_model_layout, text_config, register_layout_detector, per-layer head geometry) and support composite multimodal wrappers (Gemma 3), unmerged PEFT/LoRA wrappers, and hybrid attention stacks (Qwen3.5 / Qwen3-Next) in state controls and activation capture
• Add the system_prompt and user_prefix input controls
• Add PASTA head profiling as a steer-time fit (HeadProfile); fix scale_position and the decode-phase attention mask under transformers v5
• Add RAD prefix-reward training (rad/utils/reward_training.py) and sequence-classification heads for the Granite families
• Extend SASA to fit from contrastive pairs and rework the ITI estimator
• Left-pack batched prompts on the HF session and before state-hook assembly in _execute_generation
• Add seeding scoped per item or per dispatch (seed_scope) and counts of the tokens generated across rollouts (Output.generated_tokens, DecodingDriver.max_rollouts_per_query)
• Fix the phased, routed, and budget-forcing decoding paths across the backend interface; make SearchDriver reward_params row-scoped and fix finished-beam detection in the search Frontier
• Declare a scope (row or call) on every RUNTIME_KWARGS_SCHEMA entry, validated at steer() and read by the Inspect provider
• Add backends/vllm/environment.py, which turns the FlashInfer sampler off by default at offline engine boot, sets VLLM_HOOK_WORKER=unified for engines using the vLLM-Hook plugin, and provides serve_environment for launched servers; install vllm-hook-plugins from the steerability-interface branch
• Extend the TRL wrappers with prompt_format on the DPO/APO preference path (with a boundary warning), validation of training_args against the installed TRL config, DPO multi-loss, and resume_from_checkpoint passed to trainer.train()
• Support reasoning models by splitting reasoning on token ids when the delimiters are special tokens (utils/thinking.py); add numeric answer extraction (utils/answers.py)
• Fix canonical_value to recurse into the elements of NumPy object arrays
• Rewrite the evaluation tutorial (evaluate_steering_pipelines.md) and remove the benchmark, metric, and use-case tutorials; add the spipe concept page and reference pages for the evaluation modules, backends, spipe, utils, and the new controls
• Add the bibtex-warning hook (docs/_hooks/), make mkdocs build --strict pass, add a test over the docs nav, and rewrite AGENTS.md and README
• Move the generics and library-wrapper notebooks under algorithms/ and port the benchmark notebooks to Inspect as studies/ with task modules
• Add the system_prompt, honest-persona prompting, routing-vs-prompting, vLLM serve, and working-with-spipes notebooks; remove the TruthfulQA composite study
• Refresh the SASA (style self-steering), RAD (prefix-reward detoxification on Granite 4.1), best-of-n, and budget-forcing notebooks
• Rerun notebooks
• Add/modify tests for the Inspect provider, batching, solvers, scorers, suite, runner, and frames; spipe codec, format, store, identity, and freeze; sweeps and the runtime-kwargs schema; Gemma 3, hybrid-attention, and LoRA-composition steering; system_prompt, user_prefix, and load_checkpoint / load_lora; the DPO wrapper, TRL resume, and training-args validation; ITI, SASA, and RAD reward training; the search frontier, reward params, phased token boundaries, max rollouts, generated tokens, and padding alignment; the vLLM engine environment; the docs nav
• Bump version to 0.5.0