[UniPCMultistepScheduler] Fix final-step NaN/inf at sigma=0 and rks dtype mismatch

[deep1blue]

Summary

Fixes #14887
Fixes #14888

Root Cause & Fixes

1. Final-step NaN / inf / _LinAlgError when final_sigmas_type="zero" (UniPCMultistepScheduler returns nan with solver_type="bh1", and inf or nan with lower_order_final=False or predict_x0=False, on the final step to sigma = 0 #14887):
   Since final_sigmas_type defaults to "zero", the final step reaches sigma_t = 0, making lambda_t = inf and h = lambda_t - lambda_s0 = inf:

   • solver_type="bh1": On a 1st-order final step (D1s is None), multistep_uni_p_bh_update set pred_res = 0 and evaluated x_t_ - alpha_t * B_h * pred_res. For bh1, B_h = hh = -inf, so -inf * 0 = nan corrupted the output even though no multistep residual exists when D1s is None. Guard the B_h * pred_res subtraction behind if D1s is not None: and set x_t = x_t_ otherwise.
   • lower_order_final=False: Attempting an order >= 2 predictor update into sigma_t = 0 gives rk = (lambda_si - lambda_s0) / inf = 0 and D1s = (mi - m0) / 0 = inf (and a singular R matrix for order >= 3). Force this_order = 1 on the last step when self.config.final_sigmas_type == "zero", matching DPMSolverMultistepScheduler.
   • predict_x0=False: With sigma_t = 0 and hh = +inf, the noise-prediction branch multiplies 0 * inf = nan. Raise a ValueError in __init__ when not predict_x0 and final_sigmas_type == "zero", matching DPMSolverMultistepScheduler.

2. torch.linalg.solve dtype mismatch under torch.set_default_dtype(torch.float64) (UniPCMultistepScheduler fails in torch.linalg.solve under a float64 default dtype: the unit entry of rks takes the default dtype #14888):
   In both multistep_uni_p_bh_update and multistep_uni_c_bh_update, the first order - 1 entries of rks inherit h.dtype (float32), whereas rks.append(torch.ones((), device=device)) used the global default dtype. Under torch.set_default_dtype(torch.float64), torch.stack(rks) promoted R to float64 while b remained float32, raising RuntimeError: linalg.solve: Expected A and B to have the same dtype. Replace torch.ones((), device=device) with torch.ones_like(h).

Validation

• Added test_last_step_to_sigma_zero_is_the_data_prediction, test_predict_x0_false_requires_final_sigmas_type_sigma_min, and test_default_dtype_float64 to tests/schedulers/test_scheduler_unipc.py.
• Verified all 68 tests in tests/schedulers/test_scheduler_unipc.py pass.

Credits & Acknowledgments

• Full credit to @Nicholas022400701, who reported both UniPCMultistepScheduler returns nan with solver_type="bh1", and inf or nan with lower_order_final=False or predict_x0=False, on the final step to sigma = 0 #14887 and UniPCMultistepScheduler fails in torch.linalg.solve under a float64 default dtype: the unit entry of rks takes the default dtype #14888, diagnosed the exact root causes across solver_type="bh1", lower_order_final=False, predict_x0=False at sigma = 0, and rks under torch.set_default_dtype(torch.float64), and provided the proposed diffs and unit tests (test_last_step_to_sigma_zero_is_the_data_prediction, test_predict_x0_false_requires_final_sigmas_type_sigma_min, and test_default_dtype_float64) directly in the issue descriptions.
• Also acknowledging @DawnofGenX (Fix UniPCMultistepScheduler under a float64 default dtype #14920) for their earlier PR addressing UniPCMultistepScheduler fails in torch.linalg.solve under a float64 default dtype: the unit entry of rks takes the default dtype #14888.