✨ Add Langfuse tracing for R2R model calls

[lexenson]

Why

R2R model operations currently arrive in Langfuse without workflow-specific trace names, which makes ingestion costs and related generations difficult to group and analyze.

What changed

• Add request-scoped trace and observation contexts that merge into LiteLLM metadata.
• Give each document ingestion one stable trace ID across Hatchet boundaries, with document and task identifiers on the trace.
• Name ingestion, search, and media operations and attach r2r workflow tags.
• Keep chunk IDs and page numbers on individual generations so concurrent operations do not overwrite shared trace metadata.
• Add provider-level fallback names and the r2r tag so LiteLLM completion and embedding calls remain identifiable outside explicitly instrumented workflows.

Testing

• 32 affected unit tests passed.
• Ruff checks passed for all changed production files.
• git diff --check passed.

Deployment note

This relies on the existing LiteLLM Langfuse callback consuming trace_name, generation_name, tags, trace_id, and trace_metadata from request metadata.

[lexenson]

Closing this for a simpler solution for now with less data tracked