Add DPM++ 2M SDE Heun (RES) sampler

[Balladie]

Following up of #9444.

DPM++ 2M SDE with Heun solver type leads to the same formulation with RES, and this work adds dpmpp_2m_sde_heun sampler and removes res_multistep_ancestral that is not working on RF currently.

Checked that it produces the same results with fixed res_multistep_ancestral in #9444, with the last Euler step disabled and the noise sampler replaced to BrownianTreeNoiseSampler.

Example results from FLUX Krea:

before, res_multistep_ancestral	#9444, res_multistep_ancestral	matched, res_multistep_ancestral	dpmpp_2m_sde_heun

[comfyanonymous]

can you leave the res_multistep_ancestral_* samplers in? To ensure a clean migration I need to coordinate with the frontend people.

[Balladie]

can you leave the res_multistep_ancestral_* samplers in? To ensure a clean migration I need to coordinate with the frontend people.

restored them

[nestflow]

Thanks for the great work! Since you mentioned dpmpp_2m_sde_heun is basically the same as res_multistep_ancestral (maybe a little improved?), I wonder if we could migrate CFG++ implementation to it?

At least in my test dpmpp_2m_sde_heun does feel a bit better than res_multistep_ancestral for SDXL models, so it would be nice to have the CFG++ variation to compare with res_multistep_ancestral_cfp_pp.

[Balladie]

Thanks for the great work! Since you mentioned dpmpp_2m_sde_heun is basically the same as res_multistep_ancestral (maybe a little improved?), I wonder if we could migrate CFG++ implementation to it?

At least in my test dpmpp_2m_sde_heun does feel a bit better than res_multistep_ancestral for SDXL models, so it would be nice to have the CFG++ variation to compare with res_multistep_ancestral_cfp_pp.

please refer to #9820 @nestflow