feat: Adding support for streaming "create responses" response

[wookievx]

Brief summary of changes:

1. Introducing explicit modeling of streaming /response model, as defined by: https://developers.openai.com/api/docs/guides/streaming-responses.
2. Supporting fs2/zio/akka/ox.
3. Handling observed discrepancy from the official documentation: https://developers.openai.com/api/reference/resources/beta/subresources/responses/streaming-events#response.function_call_arguments.done (no name in actual responses, removed from the Scala model)
4. Added newest generation of models (gpt-5.6-x, to documented model list)

[flsh86]

LGTM

[adamw]

Review findings:

Bugs

1. ChatRequestBody.scala:360 — ChatCompletionModel.GPT56Terra has wire value "gpt-5.4-terra" instead of "gpt-5.6-terra" (the ResponsesModel counterpart is correct).
2. ChatRequestBody.scala:361 — ChatCompletionModel.GPT56Sol has wire value "gpt-5-6-sol" instead of "gpt-5.6-sol".
3. ResponsesResponseBody.scala:520 — unknown content-part types are not covered by the forward-compat fallback: response.content_part.added/.done are known event types, so an unmodelled part type (e.g. a future output_audio) fails decoding and errors the whole stream — the case the Unknown event/item fallbacks were added to prevent.
4. ChatRequestBody.scala:354 — none of the 8 new ChatCompletionModel case objects are in ChatCompletionModel.values, so e.g. "gpt-5.2" in a response decodes to CustomChatCompletionModel("gpt-5.2") and never matches the constant. (These chat-model additions also bypass the model_update_scripts workflow — likely how the two typos above got in.)
5. ResponsesModel.scala:102 — GPT56Terra and GPT56Sol are missing from ResponsesModel.values (only GPT56Luna was added), so they never decode from responses.

Cleanups

6. ResponsesStreamEvent.scala:450 — nothing ties the sealed-trait cases to the hand-maintained KnownTypes sets. Adding a new event but forgetting the KnownTypes entry compiles and passes tests, yet the event always decodes as Unknown in production. A wire-type → leaf-decoder Map (with KnownTypes := map.keySet), or a test enumerating the ADT cases against the set, would close this.
7. ResponsesStreamEventFixture.scala:64 — jsonCompletedWithFullUsage splices usage into jsonResponseBasic by substring surgery, producing two top-level "usage" keys; the test passes only because circe keeps the last one. deepMerge on parsed JSON would be robust.
8. ResponsesStreamEvent.scala:649 — ResponsesStreamEvent.LogProb/TopLogProb duplicate OutputContent.LogProb/TopLogProb; the same wire object surfaces as two Scala types depending on which event delivered it. OutputTextDelta/Done could reuse OutputContent.LogProb.
9. OpenAIManualCodecs.scala:301 — responsesStreamEventDispatch rewrites the JSON object and re-dispatches through the derived sum decoder on every event, including every output_text.delta. A precomputed wire-type → decoder Map would avoid the rewrite and the double dispatch (and subsume KnownTypes, see 6).
10. OpenAIResponsesOxStreamingIntegrationSpec.scala:87 — the function-tool-call integration test has no maxOutputTokens; the sibling test caps at 16. A small cap would keep the paid call cheap.

[wookievx]

Resolved the issues