Add privacy-conscious, offline-resilient analytics

[dewana-sl]

Summary

Adds dependency-free generation analytics to the Python SDK with durable, bounded delivery for offline and short-lived workloads.

Data contract

• Sends one event for each generate, generate_to_file, or completed generate_stream invocation.
• Records SDK and Python versions, OS, model and model version, selected voice, generation type, asset source, and a coarse SDK error code.
• Uses a random installation UUID and a per-event UUID.
• Never records or sends input text, generated audio, stack traces, file paths, hostnames, usernames, or project names.
• Sends only to the KittenTTS intake; no PostHog SDK or vendor dependency is added.

Offline and failure behavior

• Persists the small metadata event before starting daemon background delivery.
• KITTENTTS_OFFLINE=1 and HF_HUB_OFFLINE=1 make zero analytics network requests while keeping bounded pending events for a later retry.
• Pending storage is capped at 1,000 events and 30 days, with oldest events discarded first.
• Transient timeouts, 429, and server failures remain queued; non-retryable client errors are discarded.
• Analytics delivery never raises through or blocks the public TTS API.

Privacy and controls

• Disable per model with analytics=False or globally with KITTENTTS_ANALYTICS=0.
• Also respects HF_HUB_DISABLE_TELEMETRY=1 and DO_NOT_TRACK=1.
• Opting out removes unsent KittenTTS analytics events.
• Documentation describes the installation ID as pseudonymous rather than claiming it is untraceable.
• The intake forwards only a coarse country code and no raw IP, user agent, or city-level location after the related worker change.

Rollout dependency

Deploy dewana-sl/kittentts-analytics-worker#1 before merging this PR. It accepts stream, deduplicates retries with PostHog $insert_id, adds bounded validation and rate limiting, and returns retryable failures when upstream delivery is unavailable.

Validation

• python3 -m unittest -q (28 passing)
• python3 -m compileall -q kittentts tests
• git diff --check

The production intake currently returns 400 for generation: stream; the worker PR fixes that contract mismatch before SDK rollout.