Add OpenTelemetry tracing integration

[Tristan-Stoltz-ERC]

Summary

The cognitive loop emits tracing spans but doesn't export them via OpenTelemetry. Add optional OTLP export (behind a feature flag) so users can visualize cognitive cycle timing in Jaeger or Grafana Tempo.

Key spans to instrument

Span	Current time	% of cycle
hdc_encode	~580µs	13.6%
predict	~560µs	13.1%
consciousness_measure	—	part of postprocess
moral_algebra	~17µs	0.4%
parallel_postprocess	~1.9ms	44%
stability_regime	~1.35ms	31.6%

Total cycle: ~4.3ms (234Hz) in release mode.

Implementation plan

1. Add feature flag: opentelemetry (or otlp)
2. Add dependencies: opentelemetry, opentelemetry-otlp, tracing-opentelemetry
3. Initialize OTLP exporter in the cognitive loop constructor when feature is enabled
4. Ensure existing tracing::instrument spans are properly propagated
5. Add configuration for OTLP endpoint (default: http://localhost:4317)
6. Graceful degradation when collector is unavailable

Why this matters

Visualizing the cognitive cycle as a trace flamegraph makes it much easier to identify performance regressions, understand phase ordering, and debug consciousness measurement timing. This is especially valuable as we add more subsystems to the pipeline.

Getting started

The codebase already uses the tracing crate extensively. The main integration point is src/cognitive_loop/cycle.rs where the core pipeline runs.