test: Add token span attribute integration tests

[Pvnn]

Overview

Adds integration-style tests for token-to-span attribute assignment across all instrumented LLM providers, covering the full pipeline:

Provider utils → SpanIOProcessor-patched span.set_attribute → final span attributes

---

What's tested

Provider	Happy Path	Alt Keys	Missing Fields	Zero Values	Partial Data	Special
SpanIOProcessor	✅	—	—	—	—	aliasing
OpenAI	✅	✅	✅	✅	✅	cached + reasoning
Groq	✅	✅	✅	✅	✅	—
Cerebras	✅	—	✅	✅	✅	object vs dict details
LiteLLM	✅	✅	✅	✅	✅	reasoning tokens
DSPy	✅	✅	✅	✅	✅	—
Pydantic AI	✅	—	✅	—	✅	string cast + dual-write
Google GenAI	✅	—	✅	✅	✅	thoughts summing
OpenAI Streaming	✅	—	—	—	—	chunk accumulation

39 tests, all passing.

---

Findings during implementation

1. Zero token values silently dropped (all providers) Extraction uses usage.get("prompt_tokens") or usage.get("input_tokens") — 0 is falsy so it resolves to None before any is not None guard runs. A valid zero completion (e.g. safety-blocked response) leaves no token attributes on the span.

2. Pydantic AI casts token values to strings _safe_set_attribute calls str(value) unconditionally before writing. Token counts land as "100" instead of 100. Downstream tools doing arithmetic on token metrics may need to consider this?

---

Test approach

A patched_span fixture creates a MagicMock span, wires set_attribute to store into span.attributes, then calls SpanIOProcessor.on_start(span) to apply the aliasing closure — replicating the exact runtime pipeline with no tracer or OTLP overhead.

[akash-vijay-kv]

Either split into two focused tests (one for chunk accumulation, one for finalize → span attributes), or extract a helper that sets up the StreamingWrapper with patches.

[Pvnn]

Extracted _make_streaming_wrapper helper and split into focused accumulation and finalization tests.