Added non-openAI backends support

[lbialy]

Closes #166

BaseUri is intentionally made for user to include /v1/, check out the use of openai library in python:

from openai import OpenAI

client = OpenAI(
    base_url = 'http://localhost:11434/v1',
    api_key='ollama', # required, but unused
)

compare:

val client = OpenAISyncClient("ollama", uri"http://localhost:11434/v1")

Things left to figure out:

• Lots of APIs do not work, most of the time what does work is just /v1/chat/completions, /v1/completions and if weather, moon phase, time of day and horoscopes are right - /v1/embeddings too. Ollama doesn't implement /v1/models but Grok does and so on.
• Error modes are slightly different. When other companies / foss orgs say that API is openAI-compatible it doesn't mean it handles errors the same way. For example: ollama behaves fine and if you use a non-existent model you get a nice:

  sttp.openai.OpenAIExceptions$OpenAIException$InvalidRequestException: statusCode: 404, response: {"error":{"message":"model 'mixtral-8x7b-32768' not found, try pulling it first","type":"api_error","param":null,"code":null}}

  which is sort of what I'd expect (and what sttp-openai would expect too). For Grok however that's not as nice because their error json doesn't contain one of the expected fields, apparently:

  scala> println(oai.createChatCompletion(ChatBody(model = ChatCompletionModel.CustomChatCompletionModel("mistral"), messages = Seq(Message.UserMessage(content = Content.TextContent("What is Scala programming language?"))))).choices.head.message.content.trim)
  upickle.core.Abort: missing keys in dictionary: param

λ curl https://api.groq.com/openai/v1/chat/completions -w "%{http_code}" \
        -H 'Content-Type: application/json' \
        -d '{
          "model": "mistral",
          "messages": [
            {
              "role": "user",
              "content": "How do I shot web?"
            }
          ]
    }'
{"error":{"message":"Invalid API Key","type":"invalid_request_error","code":"invalid_api_key"}}
401

λ curl https://api.groq.com/openai/v1/chat/completions -w "%{http_code}" \
        -H 'Content-Type: application/json' \
        -H 'Authorization: Bearer REDACTED' \
        -d '{
          "model": "mistral",
          "messages": [
            {
              "role": "user",
              "content": "How do I shot web?"
            }
          ]
    }'
{"error":{"message":"The model `mistral` does not exist or you do not have access to it.","type":"invalid_request_error","code":"model_not_found"}}
404

Now this is probably another issue with upickle because param field is usually defined as param: Option[String] and when it's not present in the json is should be deserialized to None. Maybe we should configure upickle to be more permissive in error json parsing?

[lbialy]

I've fixed missing field error serialization issue in last commit. Hopefully we won't see many other error channel issues with this.

[adamw]

Thanks :)