Add rank-affine managed image reward processes

[nussejzz]

Summary

• launch one bounded scorer child per reward rank in an explicit Python environment without extra Ray GPU allocation
• add image/image-edit wire decoding, stable request chunking, identity/version checks, and deterministic idempotency
• expose resident, per-call, offload/onload, drain, and graceful shutdown lifecycles
• make EditReward support explicit model, dtype, head, batching, and device lifecycle configuration

This is PR 2 of a 3-PR stack and is based on PR 1.

Test plan

• 25 managed reward and wire contract tests
• inherited-socket real child-process smoke
• Hydra config resolution and 2,286 recipe target checks
• pre-commit checks
• top-of-stack rank-affine EditReward DP8 smoke

Made with Cursor