feat(wasm): bind task-local training losses

[GBeurier]

Summary

• bind JavaScript-local training losses directly from native DAG-ML NodeTask requirements
• return a typed TrainingLossBinding containing the local function and exact required attestation JSON
• reject PREDICT, attestation drift, descriptor mismatch, out-of-range roles, lossy numbers, and coercible non-number indices
• document controller-side ownership and extend Node 20/22 WASM smoke coverage

Stack

• depends on feat(python): execute typed local criteria #24
• continues the DAG-ML-native custom-loss runtime introduced by feat(criteria): add binding-local loss registries #23 and feat(python): execute typed local criteria #24

Validation

• cargo fmt --all --check
• cargo test -p dag-ml-wasm (5 passed)
• cargo clippy -p dag-ml-wasm --all-targets -- -D warnings
• wasm-pack build crates/dag-ml-wasm --target nodejs --out-dir target/wasm/dag-ml-wasm --release
• generated .d.ts verified as bind_training_loss(...): TrainingLossBinding with role_index: number
• independent review: no findings after two correction rounds

Local limitation

Node is not installed in the local environment, so scripts/smoke_wasm_bindings.cjs is delegated to the repository Node 20/22 CI matrix.