Fix MLite microbatch loss and forward-only output contracts

[ISEEKYAN]

Summary

• Mirror VERL’s Megatron loss-reduction hook: pass logical_loss * num_microbatches to the schedule while retaining MLite’s standard schedule-side microbatch averaging.
• Keep backward loss separate from per-microbatch reporting, preserving every original loss, model output, Metric accumulator, and plain metric through the reduction store.
• Preserve loss-context propagation across PP/VPP and PP1 forward-only token log probabilities with optional entropy.

Why

VERL PPO/SFT losses are already contributions normalized against the logical global batch. Megatron does not change its runtime API for this case: its VERL postprocess hook compensates for the schedule’s fixed microbatch averaging and reports the unscaled reduction payload separately. MLite now follows the same contract, keeping connector-specific normalization out of the public runtime interface.

Scope

This PR contains only MLite runtime/connector code and focused tests. It does not include launch scripts, training configurations, or changes to the external VERL repository.

Validation

• Focused pytest: 59 passed (test_loss_microbatch_contract, test_ops_data_trainstep_unit, test_runtime_backend_unit, test_bridge_backend, and test_mlite_engine_forward_only).
• Slurm validation job 13202007: COMPLETED, exit code 0:0.
• Local focused pytest: 17 passed.
• Ruff checks, Python compile checks, and git diff --check passed.
• Commit history, filenames, and full diff passed mechanical internal-identifier scans; the branch contains one commit on current main.