fix(litellm): preserve stream types in streaming wrappers and isolate span attribute extraction errors

[sushantsimpplr]

fixes #3577

Instrumentation must be transparent to callers of the instrumented functions. Three related defects in openinference-instrumentation-litellm break that, with litellm itself as the highest-impact caller — its Responses API → chat-completions bridge re-enters the instrumented litellm.acompletion and type-checks the result, so with this instrumentor active every streamed /v1/responses request to a provider without native Responses support fails with Unexpected response type: <class 'async_generator'>, and a secondary serialization crash masks the error (details in the issue).

Changes

1. Type-preserving stream wrappers — new _TracedSyncStream / _TracedAsyncStream (wrapt.ObjectProxy, already a dependency) present as the original stream (isinstance passes, attribute access delegates) while iterating the existing finalizing generators, so span accumulation/finalization is unchanged. The four streaming return sites (completion, acompletion, responses, aresponses) now return the proxy instead of the bare generator.
2. Exact-type gate for aresponses streaming — mirrors the sync wrapper's existing gate: streams that are not ResponsesAPIStreamingIterator pass through untouched (previously the finalizer silently drained them into an empty stream). Both responses wrappers now also end the span on the pass-through path (previously leaked).
3. Exception-isolated attribute extraction — _suppress_extractor_errors decorator on the _instrument_func_type_* extractors and _finalize_span: extraction failures (e.g. ValueError: Circular reference detected from json.dumps on self-referencing kwargs seen on litellm router retries) log and drop attributes instead of raising into the traced call.

Behavior before/after

LiteLLMInstrumentor().instrument()
resp = await litellm.acompletion(model=..., messages=..., stream=True)
# before: type(resp) is async_generator  → litellm's Responses bridge raises
# after:  isinstance(resp, CustomStreamWrapper) is True; iteration + span capture unchanged

Testing

• 4 new tests in tests/test_streaming_transparency.py: sync + async stream type preservation (content and span capture asserted), foreign-iterator pass-through on aresponses, and extractor isolation on cyclic kwargs.
• Full existing package suite passes: 115 passed. ruff check / ruff format --check clean on changed files.

Notes / shortcomings

• The anthropic-messages streaming wrappers likely share the type-swap pattern; left out to keep this PR small — happy to follow up.
• _json_serialize in openinference-instrumentation core could additionally be made cycle-safe (benefits all instrumentors); also left out of scope here.
• We currently run this exact fix in production at Simpplr as a pinned monkeypatch shim and will drop it in favor of the released fix.

[github-actions]

CLA Assistant Lite bot:
Thank you for your submission, we really appreciate it. Like many open-source projects, we ask that you sign our Contributor License Agreement before we can accept your contribution. You can sign the CLA by just posting a Pull Request Comment same as the below format.

---

I have read the CLA Document and I hereby sign the CLA

---

_{You can retrigger this bot by commenting recheck in this Pull Request}