Out of memory is not well handled

[a1exwang]

Foundry local version 0.6.87+e69a6c3d2b

Repro steps:

1. Load model A
2. Call /v1/chat/completions API with model A
3. Load model B
4. Call /v1/chat/completions API with model B

The server returns http 500 error without an error message or error code. So the client is unable to know what happened. By checking the log, I can see

E:\_work\1\s\onnxruntime\core\providers\cuda\cuda_call.cc:129 onnxruntime::CudaCall E:\_work\1\s\onnxruntime\core\providers\cuda\cuda_call.cc:121 onnxruntime::CudaCall CUDA failure 2: out of memory ; GPU=0 ; hostname=ALEX-P14S ; file=E:\_work\1\s\onnxruntime\core\providers\cuda\cuda_execution_provider.cc ; line=287 ; expr=cudaDeviceSynchronize();

Expected:

1. The 500 error body should contain error code or message indicating out of memory error.
2. The model could be automatically unloaded if not used for sometime like Ollama.

AB#74041