fix: OpenAiClient streaming error handling

[declark1]

This PR fixes an issue where OpenAI/vLLM errors sent over a stream were not being deserialized and passed through properly.

Example: previously, the below error message was being returned as a generic internal server error.

event: error
data: {"code":400,"details":"Requested sample logprobs of 21, which is greater than max allowed: 20"}

data: [DONE]

Integration test error cases will be added as part of #417.

[mdevino]

Looks good to me. Thanks for figuring this one out!