Add split GPU text encoder cache

[Jacid23]

Summary

• Add an optional split-GPU text encoder mode for systems with multiple CUDA GPUs.
• When enabled, selected text encoders are loaded on the secondary CUDA device while the main generation model stays on the primary execution device.
• Add active load/unload sync endpoints so turning the toggle off releases the secondary GPU cache instead of leaving the encoder resident.
• Add compact hardware/cache status in the UI and a model-cache sleep timer setting for idle cleanup.

Why

Text encoder loads can force the denoise model to unload/reload on single-device cache paths. On dual-GPU systems, keeping text encoders resident on the other CUDA device avoids that churn and makes repeated generation materially smoother.

Behavior

• The UI control is only useful when at least two CUDA devices are available.
• Disabling the toggle actively drops the split-GPU text encoder cache so that GPU can be used elsewhere.
• CPU offload behavior is not changed.

Verification

• pnpm lint:prettier
• pnpm lint:tsc
• pnpm lint:knip
• OpenAPI schema generated output matches checked-in openapi.json
• Typegen output is stable after regeneration

Notes

This branch was prepared from upstream/main and squashed to one focused commit. It does not include local fork/runtime update scripts, batch-specific files, or unrelated compatibility work.