feat: add OpenTelemetry distributed tracing

[Siddhant-K-code]

Motivation

Prometheus metrics (#7) cover aggregate counters and histograms. Distributed tracing adds per-request visibility — see exactly how long each pipeline stage takes and where bottlenecks are.

Closes #9

Changes

• pkg/telemetry/ — OTEL provider init with OTLP, stdout, and no-op exporters. Span helpers for each pipeline stage. W3C Trace Context + Baggage propagation. Configurable sampling rate.
• cmd/api.go — Root span for /v1/dedupe, child spans for embedding, clustering, selection, MMR. Error recording on failures. Result attributes on root span.
• cmd/serve.go — Root span for /v1/retrieve with result attributes and error recording.
• pkg/config/config.go — TelemetryConfig and TracingConfig structs, defaults, validation (exporter, sample_rate), env interpolation, updated YAML template.
• README.md — Tracing section with config examples, span table, env var usage.

Spans per request

distill.request (root)
├── distill.embedding      (if text chunks need embedding)
├── distill.clustering     (agglomerative clustering)
├── distill.selection      (representative selection)
└── distill.mmr            (MMR re-ranking, if targetK set)

Config

telemetry:
  tracing:
    enabled: true
    exporter: otlp
    endpoint: localhost:4317
    sample_rate: 1.0
    insecure: true

Testing

• 13 unit tests: disabled/enabled init, all exporters, invalid exporter, sample rate, shutdown, all span helpers, RecordResult, RecordError, zero-input edge case
• All existing tests pass
• go vet clean