perf(flux2): optimize cache locking in Klein encoder to fix #7513

[girlyoulookthebest]

Summary

Addresses a race condition in the ModelCache where models are evicted prematurely on low-VRAM systems (Issue #7513).

• Refactored invoke() to hold a single ExitStack for the entire duration of the node's execution. This ensures the Text Encoder remains locked in RAM while the conditioning data is being saved.

• Reordered the loading sequence in _encode_prompt. The Text Encoder is now locked immediately upon retrieval, closing the window where it could be evicted while the Tokenizer is being fetched.

Related Issues / Discussions

Issue #7513

QA Instructions

1. Use a system with tight VRAM where cache eviction is likely.
2. Run a generation using the Flux2 Klein text encoder.
3. In the developer log, confirm that the warning no longer appears.
4. Performance-wise, confirm that the text encoder is loaded once and stays in RAM, eliminating disk reloading churn.

Merge Plan

Safe to merge immediately. This is an architectural optimization restricted to the flux2_klein_text_encoder node and does not affect other pipelines.

Checklist

• The PR has a short but descriptive title, suitable for a changelog
• Tests added / updated (if applicable)
• ❗Changes to a redux slice have a corresponding migration
• Documentation added / updated (if applicable)
• Updated What's New copy (if doing a release after this PR)