feat: first-class OpenTelemetry export (one span per boundary)

[tishachawla-jg]

Chronicle already bootstrapped OTel and had an envelope→attributes mapper, but nothing turned a recorded run into spans. This makes OpenTelemetry first-class: each recorded boundary crossing becomes a span in Phoenix or any OTel backend.

What

with chronicle.record("run-1") as session:
    chronicle.instrument_otel()      # attaches to the active recording session
    run_agent(...)

• One OpenTelemetry span per boundary crossing, using OpenInference semantic conventions: LLM/tool span kind, input.value / output.value, llm.model_name, llm.token_count.*, tool.name, plus chronicle.* ids.
• Spans nest by the run's parent linkage, so the trace shows the agent's call graph.
• Error envelopes set span status ERROR.
• Returns an uninstrument callable.

Design

• chronicle/otel.py: instrument_otel(tracer=None, session=None) wires a new session.on_record hook that emits and ends a span per crossing; envelope_span_attributes() maps one Envelope to OpenInference attributes.
• session.py: new on_record hook fired with the full Envelope after each record (additive; default None).
• __init__: instrument_otel / envelope_span_attributes exposed lazily via module __getattr__, so import chronicle never imports opentelemetry. The base install needs neither the OTel SDK nor OpenInference; they come with the existing [phoenix] extra.

Tests (tests/test_otel.py, in-memory exporter)

One span per crossing with OpenInference attrs; parent nesting; tool name; model + token-count mapping; ERROR status on a failing boundary; uninstrument stops spans; and a subprocess test proving import chronicle does not import opentelemetry. Full suite green; ruff 0.15.22 clean. No em/en dashes.