fix(pt/pd): fix eta computation (deepmodeling#4886)

fix eta computation code

<!-- This is an auto-generated comment: release notes by coderabbit.ai
-->

## Summary by CodeRabbit

* **Bug Fixes**
* Improved ETA accuracy in training/validation progress logs by adapting
calculations to recent step intervals, reducing misleading estimates
early in runs.
* Consistent behavior across both backends, providing more reliable
remaining-time estimates without changing any public interfaces.

<!-- end of auto-generated comment: release notes by coderabbit.ai -->

Author: HydrogenSulfate

deepmd/pd/train/training.py

@@ -918,7 +918,9 @@ def log_loss_valid(_task_key="Default"):
                 self.t0 = current_time
                 if self.rank == 0 and self.timing_in_training:
                     eta = int(
-                        (self.num_steps - display_step_id) / self.disp_freq * train_time
+                        (self.num_steps - display_step_id)
+                        / min(self.disp_freq, display_step_id - self.start_step)
+                        * train_time
                     )
                     log.info(
                         format_training_message(

deepmd/pt/train/training.py

@@ -1046,7 +1046,9 @@ def log_loss_valid(_task_key="Default"):
                 self.t0 = current_time
                 if self.rank == 0 and self.timing_in_training:
                     eta = int(
-                        (self.num_steps - display_step_id) / self.disp_freq * train_time
+                        (self.num_steps - display_step_id)
+                        / min(self.disp_freq, display_step_id - self.start_step)
+                        * train_time
                     )
                     log.info(
                         format_training_message(