feat: Support bfloat16 and ensure valid precision and activation functions consistent everywhere

[njzjz]

Fix #3553.

To support bfloat16,

• DP: As NumPy doesn't plan to support bfloat16, use ml_dtypes.bfloat16. The ml_dtypes package, developed by Google, is required.
• TF: use tf.make_ndarray instead of manually extracting information from a tensor node.
• PT: As ml_dtypes is not supported yet (Allow try except check for numpy bfloat16 representation pytorch/pytorch#109873), use float32 as a bridge.

[codecov]

Codecov Report

All modified and coverable lines are covered by tests ✅

Project coverage is 77.71%. Comparing base (3c3e2ce) to head (626e35e).
Report is 3 commits behind head on devel.

Additional details and impacted files

@@           Coverage Diff           @@
##            devel    #3601   +/-   ##
=======================================
  Coverage   77.70%   77.71%           
=======================================
  Files         433      433           
  Lines       37491    37471   -20     
  Branches     1620     1620           
=======================================
- Hits        29134    29120   -14     
+ Misses       7496     7490    -6     
  Partials      861      861           

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