[MinYeFa]

Summary

Fixes #15090.

In DEISMultistepScheduler, evaluating polynomial integral coefficients in float32 leads to catastrophic cancellation as num_inference_steps increases (when the step size $h = \Delta \log\rho \ll 1$).

Specifically, in multistep_deis_second_order_update and multistep_deis_third_order_update:

• rho_t, rho_s0, rho_s1, rho_s2 were calculated via PyTorch float32 division.
• Passing float32 tensors into np.log produces single-precision NumPy floats.
• In ind_fn, the polynomial integration formulas compute finite differences between adjacent terms whose difference is $\mathcal{O}(h^2)$ or $\mathcal{O}(h^3)$ while individual terms are $\mathcal{O}(1)$.
• For 500 inference steps ($h \approx 0.002$), the denominator is $\sim 10^{-6}$ and floating-point cancellation in float32 exhausts all 24 bits of mantissa precision, leading to >8% relative trajectory error.

Solution

• Promotes rho_t, rho_s0, rho_s1, rho_s2 to standard 64-bit Python floats (float(sigma / alpha)).
• Evaluates the polynomial coefficients in full 64-bit float precision before scaling model_output_list tensors (float(...)).
• Added a dedicated regression test test_high_step_coefficient_precision verifying stable ODE integration across 50 and 100 inference steps for orders 2 and 3.