feat: Add token metrics and OpenAI-compatible response format to text generation

[nyo16]

Summary

This PR adds timing metrics and opt-in OpenAI-compatible response formats to Bumblebee's text generation serving, making it easier to integrate with
existing OpenAI-compatible tooling and monitor generation performance.
This will help bumblebee to integrated easier with other libs without writing custom middlewares .

New Options

Three new options for Bumblebee.Text.generation/4:

• :include_timing - When true, includes performance metrics in results
• :output_format - Supports :bumblebee (default), :openai, and :openai_chat
• :model_name - Model identifier for OpenAI format responses

Features

Timing Metrics

When include_timing: true:

• generation_time_us - Total generation time in microseconds
• tokens_per_second - Generation throughput
• time_to_first_token_us - Time to first token (streaming only)

Finish Reason Tracking

Tracks why generation stopped:

• "stop" - EOS token reached
• "length" - Max length reached

OpenAI-Compatible Formats

Text Completions (:openai):

%{
  id: "cmpl-...",
  object: "text_completion",
  model: "gpt2",
  choices: [%{index: 0, text: "...", finish_reason: "stop"}],
  usage: %{prompt_tokens: 5, completion_tokens: 20, total_tokens: 25}
}

Chat Completions (:openai_chat):
%{
  id: "chatcmpl-...",
  object: "chat.completion",
  model: "gpt2",
  choices: [%{index: 0, message: %{role: "assistant", content: "..."}, finish_reason: "stop"}],
  usage: %{prompt_tokens: 5, completion_tokens: 20, total_tokens: 25}
}

Both formats also support streaming with proper chunk formatting.

Backward Compatibility

All changes are opt-in with sensible defaults:
- include_timing: false
- output_format: :bumblebee

Existing code works unchanged.

Test Plan

- All existing tests pass
- Added tests for timing metrics (streaming and non-streaming)
- Added tests for OpenAI format output
- Added tests for OpenAI chat format output
- Added tests for OpenAI streaming formats
- Manually tested with Qwen3-0.6B model

[jonatanklosko]

Thanks for the PR.

To be honest, I am hesitant to add this. If we have completions compatible output, we should likely support completions input, otherwise it's incomplete. But the input is higher level than the current serving (e.g. agent/user chunks, instead of formatted prompt). Similarly, if a model supports tool calling, one may expect that we would emit tool chunks, which we don't.

Other than that, the text serving is already a bit of complex because of streaming, and adding multiple formats feels like unnecessary complexity.

Finally, while Chat Completions API are widely used, there is already Responses API, and that's only OpenAI. If we add one, it's fair to expect we may support others, and maintaining many formats doesn't seem like a good direction to me.