fix any remaining memory leaks

[ssube]

#136 should help reduce memory usage and allows turning off the cache entirely with ONNX_WEB_CACHE_MODELS=0. With that, memory usage appears to drop back to baseline between runs on Ubuntu 22, but I need to test it on Windows as well.

There are some remaining issues or limitations with the current cache:

• if the cache limit is > 0, you can still exhaust memory by loading very large models
  • the default limit of 3 and SD v2.1 models in fp32 will easily max out a 24GB card

[ssube]

The largest and most problematic leak has been fixed usptream by ORT now:

• can't start new thread error after running for a while #170
• Fix CPU memory leak due to external weights not getting memory unmapped when using non-CPU EP. microsoft/onnxruntime#15040
• fix return of flakey mnist test on gpu. microsoft/onnxruntime#1446

There are a few leaks remaining in the app, especially when using LoRAs heavily or running on DirectML where cuda.gc() doesn't do much:

• Diffusion model not fully unloading from gpu when removed from cache #242
• blend ONNX with LoRA and Textual Inversions in memory #243

[ssube]

#242 remains an issue on Windows, but is related to the DirectML runtime/drivers, and not something I can currently fix. Restarting the workers still works there.