Fix run_after names for stages with custom experiment naming

[seshakolli]

Summary

Fix cross-stage run_after dependencies for stages whose submitted experiment
names differ from NVFlow's default {expname} / {expname}-{env} convention.

WorkflowRunner previously inferred dependency experiment names directly from
stage configuration. That assumption does not hold for all stages:

• multi-environment GRPO training submits a single combined experiment such as
  {expname}-{env1}+{env2}+...
• validate_questions submits phase-specific experiments and downstream stages
  need to wait for the terminal -phase2-llm experiment

This change adds a stage-level submitted_expnames() contract so dependency
stages can report the experiment names that downstream stages should use as
run_after.

Changes

• Add BaseStage.submitted_expnames() with backward-compatible default naming
• Have WorkflowRunner query dependency stages for submitted experiment names
• Override the naming contract for:
  • GRPO training
  • validate_questions
• Document the new stage extension point
• Add regression tests covering dependency-name propagation

Testing

• 14 focused tests passed
• Full suite introduced no new failures compared with main
• ruff check passed
• ruff format --check passed
• mypy passed on changed production files
• git diff --check passed

The remaining full-suite failures observed locally are the same Windows-specific
failures present on main.