Preserve streaming timings from OpenAI-compatible providers.

[dmsnell]

llama.cpp's OpenAI-compatible streaming endpoint can include token/sec metrics in a top-level timings field on the final streamed chunk. llm was already preserving usage from that chunk, but combine_chunks() discarded provider-specific timing metadata, so llm logs --json could not show the throughput data even when the provider returned it.

In this patch, timings is captured while combining streamed chunks and include it in the persisted response JSON. Add a regression test using a llama.cpp-style streaming response.

Co-Authored-By: Codex (GPT-5.5)

[dmsnell]

I’ve seen this before, where llm would show no timing information from OpenAI-compatible API calls, but it wasn’t until working with llama.cpp locally where I examined the streaming chunks that I discovered that the information was there, but llm was discarding it.

This returns the API timing information into the logs.