[codex] Add synchronous ES LoRA trainer

[Matthew-agi]

Summary

Adds a synchronous ES-LoRA trainer as a peer trainer path under prime_rl.trainer.es, plus an es launcher/config surface.

The trainer:

• materializes LoRA candidate adapters from a flat ES parameter vector
• loads candidate adapters into existing vLLM inference pools via Prime-RL admin endpoints
• evaluates candidates synchronously across configured Verifiers train envs
• updates the mean LoRA vector with one-sided or mirrored ES estimates
• writes ES state checkpoints, the current mean adapter, metrics, and a debug config

This intentionally reuses the pieces ES needs from Prime-RL: config parsing, launch/process handling, torch distributed rank setup, env wrappers, client/admin clients, LoRA adapter serialization, logging, checkpoints, and monitors. It does not pull in SFT CE loss, optimizer, FSDP training, gradient scaling, or RL orchestrator async weight-update machinery.

Validation

Local macOS:

• uv tool run ruff check packages/prime-rl-configs/src/prime_rl/configs/es.py src/prime_rl/entrypoints/es.py src/prime_rl/trainer/es tests/unit/train/es/test_candidates.py tests/unit/test_configs.py
• uv tool run ruff format --check packages/prime-rl-configs/src/prime_rl/configs/es.py src/prime_rl/entrypoints/es.py src/prime_rl/trainer/es tests/unit/train/es/test_candidates.py tests/unit/test_configs.py
• git diff --check

Prime Intellect A6000 smoke box:

• uv run ruff check packages/prime-rl-configs/src/prime_rl/configs/es.py src/prime_rl/entrypoints/es.py src/prime_rl/trainer/es tests/unit/train/es/test_candidates.py tests/unit/test_configs.py
• uv run ruff format --check packages/prime-rl-configs/src/prime_rl/configs/es.py src/prime_rl/entrypoints/es.py src/prime_rl/trainer/es tests/unit/train/es/test_candidates.py tests/unit/test_configs.py
• uv run pytest tests/unit/train/es/test_candidates.py tests/unit/test_configs.py -q (79 passed)
• uv run es @ configs/debug/es/train.toml --dry-run
• one-step live smoke with temporary Verifiers env + vLLM PrimeIntellect/Qwen3-0.6B + dynamic LoRA loading:
  • wrote /home/ubuntu/es-smoke-output/adapter/adapter_model.safetensors
  • wrote /home/ubuntu/es-smoke-output/checkpoints/step_1/es/es_state.pt
  • metrics showed generation_s=39.07, adapter_write_s=0.22, adapter_load_s=0.19, adapter_unload_s=0.01, update_s=0.08