docs: rename q8 file refs to model_quantized.* (matches transformers.js v4 _quantized suffix for dtype=q8)

rename q8 -> quantized to match transformers.js v4 dtype="q8" filename suffix _quantized

docs: add q8 sharded variant + correct prompt format + temp guidance

Reshard q8 external data into <2GB chunks (matches onnx-community pattern)

Add q8 (MatMulNBits 8-bit, block=128) variant alongside q4f16

Re-quantize with block_size=128 + quantize Gather; reduces external data 2.83 GB -> 1.95 GB to fit browser 2 GB ceiling

Initial q4f16 ONNX export of kenpath/svara-tts-v1

initial commit