Make DeepCompile ZeRO-3 scheduling memory-aware

[tohtana]

This PR improves DeepCompile's scheduling and robustness.

• Make DeepCompile ZeRO-3 scheduling memory-aware by deriving a data-parallel-group-consistent gather budget from conservative per-node memory profiles and using it to order parameter gathers and releases.
• Preserve ZeRO parameter ownership and compiled-backward lifecycle state across eager fallback, compilation failures, and overlapping DeepSpeed engines.
• Add rank-consistency diagnostics, focused regression coverage, and documentation explaining the scheduling model and its safety constraints.