revert to working model load (first-call verified)

zerogpu: load model on cpu in parent, move to gpu per call (fixes consecutive-call crash)

readme: demo + social links, data rationale; revert to working model load; ghost dial 3/4 distinct

zerogpu: eager model load fixes consecutive-call crash; ghost dial 3/4 distinct

voice clone: transcode upload to 24k wav + auto-transcribe blank ref

fix: @gpu_task back on speak/morph (was mis-attached to DSP helper); dry cache as path not array

ZeroGPU: load F5-TTS on CPU, move to GPU inside @spaces.GPU (proper ZeroGPU pattern)

ZeroGPU: eager-load F5-TTS at module init (fix Low-level CUDA init error)

post-fx applies live: cache dry voice, re-render on dropdown change (no re-inference)

guarded @spaces.GPU decorator (no-op off ZeroGPU)

mirror app.py