Replace local ML deps with WhisperLiveKit server + Ollama-only embeddings

## Problem
The project bundles `faster-whisper`, `sentence-transformers`, and `langchain-huggingface` which pull in PyTorch and multi-GB model weights, making the Docker image and local install very heavy.

## Solution
1. Add a WhisperLiveKit Docker service for transcription (OpenAI-compatible API with speaker diarization)
2. Replace local `WhisperTranscriber` with an HTTP client calling the whisper service
3. Remove `sentence-transformers` and `langchain-huggingface` — use Ollama-only for embeddings (already the default)
4. Clean up Dockerfile (remove ffmpeg, libsndfile1)
5. Remove `faster-whisper`, `sentence-transformers`, `langchain-huggingface` from pyproject.toml deps

## Result
Python app becomes a lightweight web app with no ML dependencies. All heavy inference runs in dedicated containers (WhisperLiveKit for transcription, Ollama for embeddings).

Start with pyproject.toml and Dockerfile to inventory the listed ML dependencies and system packages, then trace the existing WhisperTranscriber. Review how the Docker services expose WhisperLiveKit and Ollama, and confirm the app uses HTTP transcription and Ollama-only embeddings. Done means the Python app runs without the removed ML dependencies and the image is lightweight.

Written by the indexing model from the issue text.