Testing all Sampler/Shedulers on Ernie-Turbo (+notes)

A comprehensive test of 93 sampler and scheduler combinations on the Ernie-Turbo model has identified the beta scheduler and seeds_3 sampler as top performers, with significant impacts on image composition at 8 steps. The beta scheduler was objectively best, while linear_quadratic was a subjective favorite for its clean details, though it may appear too clean for some subjects. Samplers like seeds_3 excelled in texture but could over-texturize faces; seeds_2 is a viable alternative. Other strong performers included exp_heun_2_x0_sde, dpmpp_2s_ancestral, dpmpp_2m_sde_gpu, and dpm_2_ancestral. Many samplers, such as dpm_fast, dpmpp_2s_ancestral_cfg_pp, and lms, produced garbage results at 8 steps. The analysis also recommends using 50 steps for base models, as 20 steps often yield poor quality.