Embedding Quantization CUDA Control

Runs one bounded 48-text BF16 control and writes the result to the attached private bucket. The requested GPU duration is capped at five minutes.

CUDA-native quantizer controls

These are bitsandbytes INT8/NF4 controls, not MLX Q/oQ/oQe replicas.

Larger-family BF16 cross-runtime controls

Larger-family CUDA-native quantizer controls