Improve save_model and load_model

[james77777778]

Fixes #19793

• save_model will first open the .h5 (in the same working dir as filepath), then write it to the zipfile at the end.
• load_model will first extract the .h5, then load it using h5py.File. The weights file will be deleted at the end.

Result:
Using GPT2 from KerasNLP

Branch	Function	Memory Usage	Time Cost
Master	save_model	1860.43MiB	3.84s
PR	save_model	1384.17MiB (-25.6%)	4.48s (+16.7%)
Master	load_model	1377.24MiB	26.93s
PR	load_model	1376.43MiB (-0.1%)	2.10s (-92.3%)

The only regression is that save_model might be slightly slower than before.

[codecov-commenter]

Codecov Report

Attention: Patch coverage is 78.72340% with 10 lines in your changes missing coverage. Please review.

Project coverage is 78.84%. Comparing base (11be99a) to head (99a8dcc).
Report is 1 commits behind head on master.

Files	Patch %	Lines
keras/src/saving/saving_lib.py	78.72%	6 Missing and 4 partials ⚠️

Additional details and impacted files

@@           Coverage Diff           @@
##           master   #19852   +/-   ##
=======================================
  Coverage   78.83%   78.84%           
=======================================
  Files         498      498           
  Lines       45864    45887   +23     
  Branches     8450     8455    +5     
=======================================
+ Hits        36159    36181   +22     
  Misses       7999     7999           
- Partials     1706     1707    +1     

Flag	Coverage Δ
keras	78.70% <78.72%> (+<0.01%)	⬆️
keras-jax	62.39% <78.72%> (+0.01%)	⬆️
keras-numpy	56.96% <61.70%> (+0.01%)	⬆️
keras-tensorflow	63.68% <78.72%> (+0.01%)	⬆️
keras-torch	62.37% <78.72%> (+0.01%)	⬆️

Flags with carried forward coverage won't be shown. Click here to find out more.

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

[Grvzard]

a tiny suggestion: the form _VARS_FNAME + ".h5" appears many times. Would it be better to define a new constant like _VARS_FNAME_H5 = _VARS_FNAME + ".h5"?

[james77777778]

a tiny suggestion: the form _VARS_FNAME + ".h5" appears many times. Would it be better to define a new constant like _VARS_FNAME_H5 = _VARS_FNAME + ".h5"?

That's cleaner. Thanks for the suggestion.
Updated.

[fchollet]

Thanks for the fix! This is great.

[fchollet]

We should put the unlink in a try/except/finally block to make sure it gets executed no matter what.

[james77777778]

I have added the try/finally blocks.
In _save_model_to_fileobj, if an error is raised, the bad model.weights.h5 won't be written to the zip file.
In _load_model_from_fileobj, if an error is raised, the extracted model.weights.h5 will be deleted no matter what.

Tests have been added to verify these behaviors.

[fchollet]

Thank you for the contribution!