Fix bug in transfer.py

[ZhengdQin]

fix the np.frombuffer in dp transfer

[codecov-commenter]

Codecov Report

Merging #1246 (84278fc) into devel (1a8fd73) will decrease coverage by 0.00%.
The diff coverage is 40.00%.

@@            Coverage Diff             @@
##            devel    #1246      +/-   ##
==========================================
- Coverage   75.99%   75.98%   -0.01%     
==========================================
  Files          91       91              
  Lines        7389     7391       +2     
==========================================
+ Hits         5615     5616       +1     
- Misses       1774     1775       +1     

Impacted Files	Coverage Δ
deepmd/entrypoints/transfer.py	72.11% <40.00%> (-0.44%)	⬇️

---

Continue to review full report at Codecov.

Legend - Click here to learn more
Δ = absolute <relative> (impact), ø = not affected, ? = missing data
Powered by Codecov. Last update 1a8fd73...84278fc. Read the comment docs.

[amcadmus]

Could you add an example to explain how to trigger the bug, what is the expected behavior and what is the actual behavior?
Thanks!

[ZhengdQin]

fix the np.frombuffer in dp transfer, the point is explained as follows:
Interpret a buffer as a 1-dimensional array, which type is the given data_dtype, which means if the type is float32, the data in the buffer will be converted to float32, no matter what the original data type is. Let's see an example:
#===========================
temp1 = np.array([1.2, 4.5], dtype=np.float64)

temp1
Out[3]: array([1.2, 4.5])

temp2 = np.frombuffer(temp1, dtype=np.float32)

temp2
Out[5]:
array([4.172325e-08, 1.900000e+00, 0.000000e+00, 2.281250e+00],
dtype=float32)
#===========================
If we do not load the data with the old type, the loaded tensor values are wrong.