Add privacy-conscious generation analytics

[dewana-sl]

Summary

Adds lightweight generation analytics to the Python SDK so maintainers can understand basic usage patterns without adding an analytics SDK dependency or collecting user text/audio.

How analytics is handled

• Sends one anonymous event for each generate, generate_to_file, or generate_stream invocation.
• Uses a local anonymous UUID persisted under the KittenTTS analytics home, not account identity.
• Captures SDK version, SDK type, platform, Python runtime, selected model, model version, selected voice, generation type, asset source, and SDK error code when a call fails.
• Does not send prompt text, generated audio, stack traces, file paths, hostnames, usernames, or project names.
• Supports opt-out with analytics=False at model construction or KITTENTTS_ANALYTICS=0.
• Sends in a daemon background thread with a short timeout and swallows analytics failures so TTS generation and process shutdown remain unaffected.

Why this is better

This keeps the SDK dependency-free for analytics. It avoids pulling in a vendor client, avoids blocking inference on network calls, and documents the event contract clearly in docs/analytics.md. Streaming is tracked once per stream invocation instead of once per chunk, which keeps event volume proportional to user actions rather than text chunking internals.

Validation

• python3 -m unittest -q
• Editable install in a fresh virtualenv with declared package dependencies.
• Import smoke for kittentts, KittenTTS, and normalize_text.
• Real inference smoke with KittenML/kitten-tts-nano-0.8-int8 and KITTENTTS_ANALYTICS=0.
• Real inference smoke with default analytics enabled, confirming generation succeeds and the process exits cleanly.