[nikich340]

Summary
Adds two endpoints to server: v1/tasks/unload_models, v1/tasks/unload_all_models.

Purpose:
In some situations, I need to free all the possible VRAM while keeping server functional and ready for new tasks without restart. That's where these features will help.

Confirmation:
Successfully built and tested on Windows 11 (CUDA sm_120 avx-512 release build) with omnivoice, qwen3-asr models. The server correctly handles new endpoints. When a model is in use, it properly waits until it finishes and then do unloading. The next requests to server work fine, with model loading back to backend device.

[0xShug0]

@nikich340 Thanks for the PR! Two minor issues:

1. LoadedModel::unload() resets model before session. Normal member destruction order would destroy session before model, and unload should preserve that safer lifetime order:
   offline=nullptr, streaming=nullptr, session.reset(), then model.reset().

2. Typo: app/server/README.md: docs say model-ids, but code requires model_ids.

Could you fix them?

[nikich340]

Sure, you are right, done!

[0xShug0]

@nikich340 Merged. Thanks!