Add Heun and DPM-Solver-2-like samplers((high-order solvers widely used in diffusion-based image generation))

[Fangjiage-1]

Summary

This PR adds two deterministic second-order sampling baselines to the PyTorch ELF implementation:

• Heun: a second-order Runge–Kutta predictor-corrector method.
• DPM-Solver-2-like: an exponential midpoint method adapted to ELF's flow-matching ODE.

Together with the existing ODE/Euler method, the sampling code now exposes three methods:

• ode
• heun
• dpm_solver_2_like

Changes

• Add _heun_step() to src/utils/sampling_utils.py.
• Add _dpm_solver_2_like_step() to src/utils/sampling_utils.py.
• Remove the previous SDE-style sampling step.
• Update src/utils/generation_utils.py to import and dispatch the three supported methods.
• Update conditional and unconditional sampling configurations.
• Keep the final sampling interval as an ODE/Euler step, following the original endpoint handling.

Sampler details

Heun

The Heun sampler uses two network evaluations per regular interval:

1. An Euler predictor estimates the next latent state.
2. A corrector averages the velocity estimates at the current and predicted states.

DPM-Solver-2-like

The DPM-Solver-2-like sampler performs:

1. An exponential half-step using the prediction at the current time.
2. A model evaluation at the midpoint.
3. An exponential full-step using the midpoint prediction.

This method is inspired by DPM-Solver-2, but it is specifically adapted to ELF's flow-matching formulation and is not a direct reproduction of the original image-diffusion implementation.

Compute matching

The provided configurations use:

Method	Sampling steps	Effective NFE
ODE/Euler	64	64
Heun	32	63
DPM-Solver-2-like	32	63

The second-order methods use two network evaluations for each regular interval, while the final interval uses the original ODE/Euler update.

Files changed

• src/utils/sampling_utils.py
• src/utils/generation_utils.py
• src/configs/sampling_configs/uncond_sampling_configs.yml
• src/configs/sampling_configs/cond_sampling_configs.yml

Validation

• Python syntax validation passed.
• Both YAML configuration files parse successfully.
• Every configured sampling method has a corresponding generation dispatch branch.
• No stale _sde_step import or dispatch remains.