Avoid redundant Megatron empty_cache and GC cleanup

[FurtherAI]

Summary

• Remove redundant gc.collect() and torch.cuda.empty_cache() calls from Megatron service, job, and step cleanup paths.
• Keep CUDA cache clearing centralized for colocated weight offload only, after weights are offloaded.
• Avoid dedicated-mode cleanup syncs between training jobs.

Tests

• Benchmarked Megatron throughput at max sequence length with CP4/EP4 Qwen3.5; completed without OOM.
• Throughput improved over previous results.