Fix double timestep shifting in FlowMatchEulerDiscreteScheduler.set_timesteps (#13243)

`__init__` records `self.sigma_max` / `self.sigma_min` *after* applying the
`shift * sigmas / (1 + (shift - 1) * sigmas)` transform. `set_timesteps` then
fed those already-shifted endpoints back into `np.linspace(...)` and applied
the same shift again, producing a different schedule for the same
`(num_train_timesteps, num_inference_steps, shift)` triple than `__init__`.

Use the pre-shift `[num_train_timesteps, ..., 1]` range directly when no
custom timesteps are provided so the shift is applied exactly once. After the
fix, `__init__` and `set_timesteps(num_train_timesteps)` produce identical
sigmas (to atol=1e-5) for every value of `shift` (verified for 1.0 and 3.0).

`self.sigma_max`/`self.sigma_min` are left as-is to preserve their public
attribute semantics for any downstream consumers.

Author: jbbqqf

src/diffusers/schedulers/scheduling_flow_match_euler_discrete.py

@@ -333,9 +333,13 @@ def set_timesteps(
 
         if sigmas is None:
             if timesteps is None:
+                # Use the raw, pre-shift `[num_train_timesteps, ..., 1]` range here:
+                # `self.sigma_max`/`self.sigma_min` are recorded *after* the shift transform
+                # in `__init__`, so feeding them back into the linspace and then re-applying
+                # the shift below would shift the schedule twice (#13243).
                 timesteps = np.linspace(
-                    self._sigma_to_t(self.sigma_max),
-                    self._sigma_to_t(self.sigma_min),
+                    self.config.num_train_timesteps,
+                    1,
                     num_inference_steps,
                 )
             sigmas = timesteps / self.config.num_train_timesteps