test: backend coverage parity audit post-refactor

[ajbozarth]

Background

Refresh of retired sub-issue #477. After the plugin refactor, verify all backends emit spans uniformly and parity holds across openai, ollama, huggingface, watsonx, litellm (and vllm if applicable).

Scope

• Add integration tests (gated on backend-specific markers, per test/MARKERS_GUIDE.md) asserting each backend produces backend-generation spans with a matching set of gen_ai.* attributes.
• Document any backend-specific attribute gaps.

Phase & dependencies

Phase 3 (polish). Depends on #1045 and #1046. Can start any time after Phase 1 lands; best done after most of Phase 2 has merged so the audit is against steady-state tracing.

Acceptance criteria

• All backends produce backend-generation spans with a shared minimum attribute set.
• Known gaps documented.

Parent epic: #444