fix(pt): make 'find_' to be float in get data (deepmodeling#3992)

make 'find_' to be float in get data, fix deepmodeling#3991 .

On my device, the profiler indicates that `cudaStreamSynchronize` takes
negligible time, resulting in minimal speedup.

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

## Summary by CodeRabbit

- **New Features**
- Enhanced data loading by adding a `collate_fn` parameter for more
flexible data collation.
- Improved data filtering by excluding keys containing "find_" in
addition to existing filters.

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

deepmd/pt/train/training.py

@@ -184,6 +184,7 @@ def get_dataloader_and_buffer(_data, _params):
                     if dist.is_available()
                     else 0,  # setting to 0 diverges the behavior of its iterator; should be >=1
                     drop_last=False,
+                    collate_fn=lambda batch: batch,  # prevent extra conversion
                     pin_memory=True,
                 )
                 with torch.device("cpu"):
@@ -1093,7 +1094,7 @@ def get_data(self, is_train=True, task_key="Default"):
                     batch_data = next(iter(self.validation_data[task_key]))
 
         for key in batch_data.keys():
-            if key == "sid" or key == "fid" or key == "box":
+            if key == "sid" or key == "fid" or key == "box" or "find_" in key:
                 continue
             elif not isinstance(batch_data[key], list):
                 if batch_data[key] is not None: