docs: add native loss and metric roadmap

[GBeurier]

Summary

• establish DAG-ML native-first ownership for loss and metric semantics
• require process-local custom loss execution in Python, R, JavaScript/WASM, Rust, C ABI, and MATLAB
• require each binding to invoke its own language-native callback during real training, not delegate to Python or store opaque metadata
• separate training loss, early stopping, selection, reporting, tuning, threshold, and ensemble roles
• define L0-L7 delivery batches, review boundaries, evidence, and test gates
• reserve score-provider-owned metric contracts and gate fingerprints on the published training/TCV1 canonicalization API

Validation

• sphinx-build -W --keep-going -b html docs docs/_build/html
• git diff --check

Independent review

• reviewed head: cfdc9732d577133cad07c8a3cb6ac2f0a3416198
• scope: roadmap and ADR-22
• verified local callback execution in every official language, metric/provider ownership, and TCV1 dependency gating
• verdict: approved with no findings

Concurrency

This PR contains documentation only. It deliberately avoids all source and shared-schema files touched by the active score-provider and training/TCV1 work.