Add Timm image classification via pluggable training backends

[whittenator]

Introduce a generalized, registry-based training-backend abstraction so
training libraries are no longer hard-coded into the orchestration code, and
add Timm (PyTorch Image Models) as a first-class image-classification trainer
alongside Ultralytics/YOLO.

Backend

• New app/services/training package: Trainer ABC + TrainContext/TrainResult,
  a registry (get_trainer/list_frameworks, defaults to ultralytics), shared
  dataset exporters (ImageFolder + YOLO-detect), and two trainers.
• UltralyticsTrainer: behavior-preserving extraction of the existing YOLO
  training/export/predict logic behind the interface.
• TimmTrainer: full classification loop with the complete Timm surface —
  any backbone (searchable), pooling head + drop rates, full optimizer/
  scheduler matrix, mixup/cutmix, model EMA, label smoothing, AMP,
  channels-last, and the full augmentation pipeline; self-describing
  checkpoint; ONNX export and predictor.
• Thin train_task / onnx_export / evaluation / inference now resolve the
  framework via the registry and delegate (no library names in orchestration).
• framework column on ExperimentRun + ModelArtifact; migration 0008 merges the
  two open migration heads and adds the columns.
• Capabilities API (/api/training/frameworks + model search) drives the form;
  TrainRequest.framework added and validated against the trainer's tasks.
• requirements: add timm and matplotlib (propagate to agent images).

Frontend

• experiments/new.tsx is now schema-driven: a Framework selector, task/model/
  hyperparameter sections rendered from the capabilities API, and a searchable
  Timm backbone picker. Ultralytics form is unchanged.

Tests

• Registry, shared dataset exporters, capabilities API, framework persistence/
  validation, Timm capability schema, and a gated CPU smoke train.

https://claude.ai/code/session_01CGFLcnxLvPEuF1y6W79nfA