fix: Shift Model Loading from Lazy Loading to Preloading with Warm-Up

Summary

This MR updates the model initialization strategy by replacing lazy loading with eager preloading at application startup, along with a dummy inference (warm-up pass) to ensure all models are fully initialized before handling real requests.

Changes Made

• Moved model loading from on-demand (lazy) to startup phase

• Implemented centralized model loader

• Added warm-up step using dummy input

• Ensures all pipelines are initialized:

  • ASR (Swecha + Whisper)
  • Language Identification (LID)
  • Diarization
  • Punctuation
  • Alignment

• Updated logs to reflect loading and warm-up stages

• Ensured models are cached and reused across requests

Motivation

Previously:

• First request experienced high latency (cold start)
• Models loaded during request → poor user experience
• Risk of race conditions under concurrent requests

Now:

• Models are ready before serving traffic
• Eliminates cold-start delays
• Improves consistency and reliability

Impact

Improvements

• Faster response time for first request
• Consistent latency across all requests
• Early detection of model loading failures
• Better production readiness

Trade-offs

• Increased application startup time (~100s observed)
• Higher initial CPU/GPU usage during boot
• Slightly heavier memory footprint at idle

---

Performance Observations

• Model loading time: ~100 seconds
• Warm-up time: ~14 seconds
• Runtime latency improved by eliminating cold start delays

---

*closes #22 (closed)