feat: add DPM++ 2M SDE sampler

[fszontagh]

Summary

Adds dpm++2m_sde - DPM-Solver++(2M) SDE, midpoint variant. It uses the same multistep skeleton as the existing dpm++2m (one model eval per step) plus a stochastic term controlled by --eta (default 1.0, matching k-diffusion). At the same seed it produces different but plausible compositions from the ODE variant, which is the expected signature of a working SDE sampler.

Reference: Lu et al. arXiv:2211.01095; k-diffusion sample_dpmpp_2m_sde.

Wired through everywhere the existing samplers are: enum, name table, CLI --sampling-method and --eta help, sd-api A1111 aliases (dpm++ 2m sde, k_dpmpp_2m_sde), and resolve_eta default.

Related Issue / Discussion

None.

Additional Information

Verified end to end on:

• SD 1.5 (v1-5-pruned-emaonly-fp16) at 512x512, 20 steps, karras, seed 42
• SDXL (AcornXL_V2) at 1024x1024, 25 steps, karras, cfg 6.0, seed 42

Same prompt ("a cat"), same seed. SDE and ODE outputs diverge as expected while remaining coherent and artifact-free. Runtime is on par with dpm++2m (SDXL 25 steps: 47.76s SDE vs 46.70s ODE).

SD 1.5, 512x512, 20 steps

dpm++2m_sde	dpm++2m (ODE)

SDXL, 1024x1024, 25 steps

dpm++2m_sde	dpm++2m (ODE)

Design notes

• Midpoint variant only. k-diffusion also exposes a heun variant, but midpoint is the community default and adding a flag for the other would just be a knob.
• randn_like per step, matching sample_er_sde and sample_res_multistep. No BrownianTree sampler added.
• No flow-denoiser branch, matching the existing sample_dpmpp_2m ODE sibling. Users on flow models should keep using euler_flow or res_multistep.
• Enum value is appended, so existing integer values stay stable.

Checklist

• I have read and confirmed this PR follows the contribution guidelines.