remove torchao dependency, quantize entirely via linear

Fix issues with loading F8Linear from state dict when init_scale not initialized & loaded from meta device

Small fix for issue where f16 CublasLinear layers weren't being used even when available.

Ensure repo only accesses CublasLinear lazily

Remove f8 flux, instead configure at load, improved quality & corrected configs

Dynamic swap with cublas linear / optional improved precision with vram drawback

Remove more unnecessary code, fix small typing hickup

Remove unnecessary code, hide prints behind debug flag, hide warnings

Add fields to configs, fix issue with offload from bnb, remove extra random text code

Fix for nightly

cuda version checks

Fix non-offload inference & add option to load from prequantized flux

Add offloading & improved fp8 inference.