pt: make get_data non-blocking

[njzjz]

to(DEVICE) is cpu-blocking but to(DEVICE, non-blocking=True) is not blocking. This improves performance by at least 0.1s/100 steps.

Before, get_data is blocking:

After, get_data is not blocking:

The subsequent blocking is phys2inter (via torch.linalg.inv).

[codecov]

Codecov Report

Attention: Patch coverage is 50.00000% with 1 lines in your changes are missing coverage. Please review.

Project coverage is 76.95%. Comparing base (b0171ce) to head (f1e1ba6).
Report is 2 commits behind head on devel.

Files	Patch %	Lines
deepmd/pt/train/training.py	50.00%	1 Missing ⚠️

Additional details and impacted files

@@           Coverage Diff           @@
##            devel    #3422   +/-   ##
=======================================
  Coverage   76.94%   76.95%           
=======================================
  Files         423      423           
  Lines       35878    35883    +5     
  Branches     1620     1620           
=======================================
+ Hits        27607    27612    +5     
  Misses       7403     7403           
  Partials      868      868           

☔ View full report in Codecov by Sentry.
📢 Have feedback on the report? Share it here.