Fix llama.cpp pin: clone tag b9553 instead of short commit ref

Fix MTP: longer startup timeout, -ngld flag, pin llama.cpp b9553

Fix MTP detection in scripts/ensure_llama_colab.py

Auto-rebuild llama-server >= b9549 for Gemma 4 MTP

Add Pass 0 speaker diarization (age/gender) to Colab + secret-free token loading

Clarify notebook docs: SKIP_CORRECTION default True, pipeline table

Remove build cell; restore llama-server from Drive tgz only

Fix llama-server build on Colab T4: nvcc, CUDA env, verbose errors

Initial upload: Gemma SRT Translate + Colab notebook