Add NAVA fine-tuning backend

[AlexEisie]

Summary

• Add a task-backend registry and move LM/Megatron dispatch behind backend/engine hooks.
• Add a self-contained NAVA task package with label conversion, LoRA patching, learner loop, export support, and layout metadata validation.
• Extend the syncer/protocol to carry and checkpoint layout metadata.
• Add NAVA G4/L4 smoke scripts with a no-text-encoder/no-VAE mock latent path for fast two-learner flow validation.

Validation

• git diff --check && git diff --cached --check
• PYTHONDONTWRITEBYTECODE=1 python -m compileall -q yeto
• cargo test --manifest-path syncer/Cargo.toml
• PYTHONDONTWRITEBYTECODE=1 PYTHONPATH=. pytest -q tests/test_backends.py tests/test_nava_launcher.py tests/test_nava_scripts.py tests/test_nava_labels.py tests/test_fragments.py tests/test_layout_metadata.py
• PYTHONDONTWRITEBYTECODE=1 PYTHONPATH=. pytest -q tests/test_controller.py tests/test_launch_auto.py tests/test_head_mode.py -k 'not local_syncer_process_restarts_from_checkpoint'
• PYTHONDONTWRITEBYTECODE=1 PYTHONPATH=. pytest -q tests/test_delivery.py tests/test_datasource.py tests/test_gpu_spec.py -k 'not upload_sky'

Live smoke

• Ran two g4dn.xlarge spot learners in us-west-2 using the mock latent NAVA smoke path.
• Verified all 4 syncer outer steps had both learners as responders.
• Verified server-side learner adapters, learner state, checkpoint/tape, and export artifacts without downloading outputs locally.
• Cleaned up the G4 spot instances, spot requests, test key pair, and security group.

Notes

• docs/NAVA_L4_SMOKE.md was intentionally excluded from this commit/PR.