refactor: auto offload strategy, add vision_quant to PipelineConfig

Remove cpu_offload, layer_offload, adaptive_offload and offload_compression
from QuantFuncPipelineConfig. libquantfunc now picks the VRAM/offload
strategy at runtime based on free VRAM and model size, which avoids users
having to hand-tune parameter combos per machine.

Legacy workflows that still set these keys continue to load — libquantfunc
silently ignores them with a one-time deprecation warning.

Add vision_quant (int8/int4/fp8/fp4/fp16, default int8) to mirror the
existing text_precision and expose vision-encoder quantization.

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>