Implement dynamic GPU model attribution (attach on demand, detach after training)

[guillaume-byte]

Summary

Implement dynamic GPU model attribution so models are attached to GPU only when needed and detached when no longer needed.

Scope

• At init/resume lifecycle, avoid unnecessary eager GPU attribution.
• Attach model to GPU on first required operation (train/resume/evaluate as applicable).
• Detach/offload model from GPU when training completes.
• Re-attach cleanly for evaluate/inference flows.

Acceptance Criteria

• Model is not unnecessarily GPU-bound at startup.
• Train/eval transitions correctly attach/detach model with no regressions.
• Resume/checkpoint flows remain functional.
• Memory behavior improvement is observable on representative runs.