Add LangSmith tracing and system-of-record webhook

[snowsecure]

Motivation

• Enable structured tracing and sampling for LLM calls using LangSmith so API calls and responses can be observed and correlated across runs.
• Provide an optional system-of-record webhook to notify downstream systems when an abstract finishes processing, enabling eventual integration with external systems of record.

Description

• Add a new LangSmith helper module app.core.langsmith and settings in app.core.config (LANGSMITH_*), plus environment examples in .env.example and docs in backend/README.md to configure tracing and sampling.
• Wire LangSmith initialization into FastAPI startup (app.main) and Celery startup (app.workers.celery_app) via configure_langsmith().
• Instrument LLM clients (app.core.gemini_client, app.core.clau de_client, app.core.azure_openai_client) to start/end sampled LangSmith runs and attach input/output previews and token metadata using build_llm_inputs, build_llm_outputs, and start_langsmith_run.
• Add an optional system-of-record publisher app.core.system_of_record and call publish_abstract_completed() from Celery processing tasks in app.workers.tasks after abstracts are saved; also add corresponding config flags and .env.example entries, and add langsmith to backend/requirements.txt.

Testing

• No automated tests were executed as part of this change.
• Existing test suite (pytest) was not run during this rollout; updates are limited to wiring/tracing and webhook publish calls and should be covered by integration/smoke tests in downstream validation.

---

Codex Task