fix(models): accumulate streamed logprobs in place to avoid O(n^2) copies

[anxkhn]

When Chat Completions streaming is used with logprobs enabled, the stream handler
rebuilt the entire accumulated output-text logprobs list on every content delta:

existing_logprobs = state.text_content_index_and_output[1].logprobs or []
state.text_content_index_and_output[1].logprobs = existing_logprobs + output_logprobs

list + list allocates and copies a brand-new list of the full accumulated length
for each delta, so over an n-token stream this is O(n^2) allocation and copying
(the adjacent ... .text += delta.content on the line above already grows in
place). The accumulator was introduced in #2134 using + where an in-place extend
was intended.

This extends the existing list in place instead, keeping the None-provider guard
so non-OpenAI providers (for example LiteLLM), which may leave logprobs unset,
still behave exactly as before. The accumulated contents and their order are
unchanged; only the per-delta work drops from O(n) to O(1), making the whole stream
O(n).

The converted logprobs are a fresh, unaliased list[Logprob] from
ChatCmplHelpers.convert_logprobs_for_output_text, and the response part's
logprobs is process-local (initialized to [] at part creation), so extending it
in place is safe and observably identical to the previous concatenation.

Test plan

• Added test_stream_response_accumulates_logprobs_across_many_deltas in
  tests/models/test_openai_chatcompletions_stream.py: streams several
  logprob-bearing content deltas and asserts the final output-text part accumulates
  every token's logprob in order (guarding that the in-place refactor preserves
  ordered accumulation). This complements the existing
  test_stream_response_includes_logprobs.
• make format (no changes), make lint, and make typecheck (pyright clean) all
  pass on the change.
• uv run pytest tests/models/ passes (all model tests, including both streaming
  logprob tests).

Issue number

Checks

• I've added new tests, if relevant
• I've run .agents/skills/code-change-verification/scripts/run.sh
• I've confirmed all verification steps pass
• If using Codex, I've run /review before submitting this PR