refactor: move tracing onto plugin/hook pattern

[ajbozarth]

Background

Tracing is the only telemetry pillar still using pre-plugin instrumentation. Metrics (#443) and logging (#442) both migrated to the canonical plugin/hook pattern during their epics; tracing needs the same treatment to unblock the rest of epic #444.

Current state

• mellea/telemetry/backend_instrumentation.py exposes instrument_generate_from_context, instrument_generate_from_raw, start_generate_span, record_token_usage, record_response_metadata.
• Each of 5 backends (openai, ollama, huggingface, litellm, watsonx) calls these helpers inline (~20 call sites total).
• The active span is stored in mot._meta["_telemetry_span"] so it can survive the asyncio.create_task boundary and be closed in post_processing() / error path.
• Tracing reads raw provider responses for attribute extraction instead of mot.generation.

Scope

• Create mellea/telemetry/tracing_plugins.py following mellea/telemetry/metrics_plugins.py as the reference. Example structure:
  • BackendTracingPlugin (priority ~50) — hooks generation_post_call in FIRE_AND_FORGET mode, reads payload.model_output.generation, emits the backend generation span with gen_ai.* attributes.
  • Register plugins through a _TRACING_PLUGIN_CLASSES tuple inside tracing.py at import time when tracing is enabled.
• Remove all direct tracing imports from mellea/backends/*.py. Delete the ~20 inline instrumentation calls.
• Remove the mot._meta["_telemetry_span"] stash. Spans now live inside the plugin and don't need to round-trip through the MOT.
  • This closes out ModelOutputThunk structural cleanup — ._meta partitioning, raw responses, _thinking #909 item 2 as a side effect.
• Keep mellea/telemetry/tracing.py as the low-level setup module (tracer provider, exporters, Tracer creation). Plugins consume get_application_tracer() / get_backend_tracer().
• Application-level spans (session, aact) stay as a separate question — tracked in feat: session and act spans via plugin hooks #1048.
• Must preserve all attributes PR feat(telemetry): close five OTel GenAI semantic convention emission gaps (#1035) #1036 lands: gen_ai.provider.name, gen_ai.conversation.id, error.type, and cache/reasoning token attrs (gen_ai.usage.cache_read.input_tokens, gen_ai.usage.cache_creation.input_tokens, gen_ai.usage.reasoning.output_tokens). Prompt-template attrs, request params, and content capture are tracked as follow-up issues (feat(telemetry): emit llm.prompt_template.* from tracing plugin (post-#1045) #1067, feat(telemetry): emit gen_ai.request.* from tracing plugin (post-#1045) #1068, feat(telemetry): emit gen_ai.input/output.messages content events from tracing plugin (post-#1045) #1073) and are not in scope here.

Phase & dependencies

Phase 1 (foundation). No internal deps. Blocks all of Phase 2 and Phase 3. Sequencing within Phase 1: this issue → #1046 → #1047; all three likely land as a single PR per the epic plan.

Coordination notes

• Streaming generation path has an open question about when GENERATION_POST_CALL fires (see feat: streaming validation — per-chunk requirement checking with early exit #891 discussion, ModelOutputThunk structural cleanup — ._meta partitioning, raw responses, _thinking #909 comment thread). Whatever resolves that concern for TokenMetricsPlugin applies here identically.
• Coordinate with @planetf1 on PR feat(telemetry): close five OTel GenAI semantic convention emission gaps (#1035) #1036 (emission gaps) — that PR is compatible to merge as-is; after it lands, the attribute emission moves from backend_instrumentation.py into the plugin added here.
• Three Phase 2 issues (feat(telemetry): emit llm.prompt_template.* from tracing plugin (post-#1045) #1067, feat(telemetry): emit gen_ai.request.* from tracing plugin (post-#1045) #1068, feat(telemetry): emit gen_ai.input/output.messages content events from tracing plugin (post-#1045) #1073) explicitly depend on this refactor to make their plugin-layer implementations possible. The plugin architecture here should accommodate a GENERATION_PRE_CALL hook (needed by feat(telemetry): emit gen_ai.request.* from tracing plugin (post-#1045) #1068) and formatter-layer attribute emission (needed by feat(telemetry): emit llm.prompt_template.* from tracing plugin (post-#1045) #1067) so those follow-ups don't hit design blockers.

Acceptance criteria

• mellea/telemetry/tracing_plugins.py exists with at least BackendTracingPlugin.
• No file under mellea/backends/ imports from mellea.telemetry.backend_instrumentation or mellea.telemetry.tracing.
• mot._meta["_telemetry_span"] is not set or read anywhere (closes ModelOutputThunk structural cleanup — ._meta partitioning, raw responses, _thinking #909 item 2).
• All attributes from PR feat(telemetry): close five OTel GenAI semantic convention emission gaps (#1035) #1036 continue to be emitted by the plugin with no regression.
• Existing tracing tests pass (adjust as needed).
• A basic integration test asserts spans are emitted for at least one backend with the new plugin path.

Parent epic: #444