feat: add DPM++ 2M SDE (Brownian tree) sampler

[fszontagh]

Summary

Adds dpm++2m_sde_bt - the DPM++ 2M SDE variant that k-diffusion / ComfyUI expose as sample_dpmpp_2m_sde_gpu. Same math as the dpm++2m_sde merged in #1742, but the noise per step comes from a seeded Brownian tree instead of a fresh randn_like, giving step-count-stable composition: the same seed at 20 and 40 steps produces the same image, only fine detail differs.

The k-diffusion _gpu suffix is legacy naming (torchsde used to require CUDA tensors) - the algorithm is scalar-side and backend-agnostic. Named _bt here for honesty, with k_dpmpp_2m_sde_gpu / dpm++ 2m sde gpu aliased in the sdapi mapper for A1111 / Comfy compatibility.

Related Issue / Discussion

Follow-up to #1742.

Additional Information

Verified end to end on SD 1.5 (v1-5-pruned-emaonly-fp16), 512x512, karras, seed 42, prompt "a cat". Same seed, only step count varies. Runtime overhead of the Brownian tree is ~1% (SD 1.5 20-step: 9.01s vs 8.93s for the non-BT variant).

Step-count stability (this is the point)

	20 steps	40 steps
dpm++2m_sde_bt (this PR)
dpm++2m_sde (baseline)

Top row: same tortoiseshell portrait, same eyes, same background - refined detail only. Bottom row: two unrelated cats.

Design notes

• BrownianTreeNoiseSampler is a recursive Brownian bridge with fixed max depth (24 levels). Each interval seed derives from the base seed via splitmix64 + branch salt, so the noise at any point is a pure function of the seed and the position within [sigma_min, sigma_max], independent of which schedule sigmas are queried. This is what gives the step-count stability.
• Base seed is drawn from the sampler RNG at loop entry, so it inherits the user seed and stays deterministic.
• W(sigma) values are memoised across the sampling loop; each schedule point is reused as sigma_a on the next step.
• Enum value is appended; existing integer values stay stable.

Checklist

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

[leejet]

Thank you for your contribution.