High VRAM usage after img2img when worker stays alive

[HoopyFreud]

Unless the worker is killed after completing an img2img job, onnx_server continues to use high VRAM, resulting in large slowdowns in subsequent generation, even when cache limit is set to zero and the server cache is empty. Can be worked around by setting ONNX_WEB_JOB_LIMIT=1

[HoopyFreud]

System state with ONNX_WEB_JOB_LIMIT=10 (default) - garbage collection has been run and no leaking worker has been detected despite VRAM usage remaining high

[HoopyFreud]

On subsequent keyboard interrupt, job failure is set for the job that just finished

[HoopyFreud]

System state with ONNX_WEB_JOB_LIMIT = 1 - worker is detected as leaking and is shut down immediately after finishing

[HoopyFreud]

This appears to have magically fixed itself with the new pipeline model