[usability] Improve SkyRL cold start time

[erictang000]

Currently SkyRL load times are dominated by kernel/cuda graph compilation. These can be optimized/reused when available + when using the same engine/trainer configs instead of compiling from scratch on existing clusters.

Some techniques for this are detailed in the Nemotron-Ultra tech report (https://research.nvidia.com/labs/nemotron/files/NVIDIA-Nemotron-3-Ultra-Technical-Report.pdf)

• [P0] Persistent warm cache on shared storage + node local seeding on startup (for fully async/non-colocated, need to make sure that inference + trainer nodes are consistently seeded to the same ips if the cluster remains the same)
• [P1] Use persistent warm cache + RDT to quickly + dynamically add new inference engines during training

[JiangJiaWei1103]

Take, please assign this to me. Thx!