Separate pipeline freshness from runtime tuning

[FurtherAI]

Summary

• move max_steps_off_policy from PipelineRuntimeConfig to a fixed PipelineTrainer keyword argument
• keep strict and mean policy freshness constraints configurable while pipeline autotuning is enabled
• update internal callsites to use the trainer-level freshness argument

Why

PipelineRuntimeConfig is rejected when autotuning is enabled because the autotuner owns its worker, batch, and queue settings. Policy freshness is a training constraint rather than a tunable runtime setting, so placing max_steps_off_policy in that object unnecessarily prevented users from configuring it with autotuning.

Mean-only in-flight update runs can now use:

PipelineTrainer(
    ...,
    autotune=PipelineAutotuneConfig(mode="online"),
    max_steps_off_policy=None,
    limit_mean_steps_off_policy=2.0,
)

The strict default remains 4.

Validation

• 72 pipeline tests passed, 1 GPU test skipped
• repository prek hooks passed: Ruff, Ruff format, ty, and lockfile sync