Fix partial-load device recovery across CLIP, T5, and Qwen text encoders

[JPPhoto]

Summary

Fix partial-load device recovery across CLIP, T5, and Qwen text encoders.

This PR broadens the partial-load recovery work in #8959 so interrupted or partially offloaded text encoder runs do not leave later executions with mixed-device state or with inputs sent to the wrong device.

The earlier repair covered Qwen/GLM-specific paths, but other supported text encoder architectures still had similar failure modes. CLIP- and T5-based paths could still rely on model.device or first-parameter device inspection, which is unreliable when a model is partially loaded. The shared loaded-model path could also leak a cache lock if repair failed during __enter__().

Related Issues / Discussions

• Follow-up to Fix/model cache Qwen/CogView4 cancel repair #8959
• User-submitted reports:
  • https://discord.com/channels/1020123559063990373/1489707504865771590
  • https://discord.com/channels/1020123559063990373/1149510134058471514/1491686043844476969

QA Instructions

Run focused backend regression tests:

• pytest tests/app/invocations/test_compel.py
• pytest tests/app/invocations/test_sd3_text_encoder.py
• pytest tests/backend/flux/modules/test_conditioner.py
• pytest tests/backend/model_manager/load/test_loaded_model.py
• pytest tests/app/invocations/test_cogview4_text_encoder.py
• pytest tests/backend/model_manager/load/model_cache/cached_model/test_repair_required_tensors.py

A practical way to test this is to run a generation with an affected model family like SDXL, SD3, or FLUX under enough VRAM pressure that partial offloading is likely, then cancel the job while it is still running and immediately start another generation with the same model. Before this fix, follow-up runs could fail with "Expected all tensors to be on the same device" because some text encoder tensors were left on CPU while execution resumed on cuda:0. To make the reproduction more reliable, cancel several runs back-to-back or queue the next run immediately after stopping the first. After this fix, the next run should recover cleanly without mixed-device text encoder errors.

Merge Plan

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)

[lstein]

Works as advertised.