feat: Add text-only inference with CPU extractor patches (Phase 0)

Fix: Ensure timeline field on all events after TRIBE v2 transforms

Fix: Manually ensure all required event columns exist by iterating through events after TRIBE v2 transforms

Fix: Use standardize_events() to validate events after TRIBE v2 transforms

Fix: Post-process events DataFrame to ensure all required columns exist after TRIBE v2 transforms

Fix: Add required columns (timeline, subject, offset, frequency, extra) to event DataFrames to pass TRIBE v2 validation

Refactor: Add multimodal support (video/audio/text) with proper TRIBE v2 pipeline

Bypass get_events_dataframe entirely, create events directly for video-only inference

Completely bypass whisperx by patching _run method to skip audio extraction

Bypass whisperx audio transcription to get video-only inference working

Fix: patch tribev2 source file to use int8 compute_type for whisperx on CPU

Fix: create uvx wrapper to force int8 compute_type for whisperx on CPU

Fix: patch whisperx to use int8 compute_type on CPU (float16 not supported)

Fix: pre-download whisperx model during Docker build to avoid 429 rate limits at runtime

Fix: pre-download whisperx model and set HF_HUB_CACHE to avoid 429 rate limits in uvx subprocess

SGP-Tribe3 v1.0 initial deploy