decouple trainer step completion from weight broadcast

[tim0120]

Summary

• publish a generic per-run trainer-step completion marker after each optimizer update
• make static SFT wait on that marker instead of a filesystem weight broadcast
• restrict filesystem and NCCL weight publication to ready runs that actually need inference
• move ready_to_update release out of broadcaster implementations and into trainer-step completion

This keeps WeightWatcher weight-specific and avoids full-model or adapter serialization for static SFT. Policy-backed RL runs continue to use their existing weight broadcast paths.

Multi-run behavior

A trainer step may update static and policy-backed runs together. Only policy-backed runs are included in weight publication; all updated runs receive their own trainer_steps/step_N/STABLE completion marker and are released for their next batch.

Validation

• focused trainer state test: 1 passed
• Ruff on all changed Python files: passed
• diff whitespace check: passed

The repository lock currently supports Linux only, so the focused test was run on macOS with an ephemeral uv --no-project dependency environment.