pt: support multitask finetune

[iProzd]

This PR：

1. merge change_energy_bias into compute_output_stats and reformat it into change_out_bias of model level.
2. support single-task/multi-task finetuning from single-task/multi-task pretrained model.

Need fix in future PR:

1. Finetuned model has covered type_map. (If fixed, change_out_bias func will not need the input params origin_type_map and full_type_map.) See also [BUG] Finetuned model has wrong type_map #3455.
2. change_out_bias support for other models.(e.g. Spin, ZBL, Polar, Dipole and Dos.)

[codecov]

Codecov Report

Attention: Patch coverage is 91.81034% with 19 lines in your changes are missing coverage. Please review.

Project coverage is 77.59%. Comparing base (145f501) to head (19b9e79).

Files	Patch %	Lines
deepmd/dpmodel/atomic_model/linear_atomic_model.py	40.00%	3 Missing ⚠️
...eepmd/pt/model/atomic_model/linear_atomic_model.py	40.00%	3 Missing ⚠️
deepmd/pt/utils/finetune.py	95.65%	3 Missing ⚠️
deepmd/dpmodel/atomic_model/dp_atomic_model.py	50.00%	2 Missing ⚠️
...eepmd/dpmodel/atomic_model/pairtab_atomic_model.py	50.00%	2 Missing ⚠️
...epmd/pt/model/atomic_model/pairtab_atomic_model.py	50.00%	2 Missing ⚠️
deepmd/utils/finetune.py	50.00%	2 Missing ⚠️
deepmd/pt/model/atomic_model/base_atomic_model.py	96.87%	1 Missing ⚠️
deepmd/pt/utils/auto_batch_size.py	96.55%	1 Missing ⚠️

Additional details and impacted files

@@            Coverage Diff             @@
##            devel    #3480      +/-   ##
==========================================
+ Coverage   77.48%   77.59%   +0.10%     
==========================================
  Files         432      432              
  Lines       37182    37314     +132     
  Branches     1620     1620              
==========================================
+ Hits        28812    28955     +143     
+ Misses       7502     7491      -11     
  Partials      868      868              

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

[wanghan-iapcm]

• It seems that the change output bias should be place in atomic model (current bias can also be computed by the forward_atomic method)
• the implementation should be able to reused by other atomic models, including the zbl and linear models, thus it is either placed in base_atomic_model, or be designed as a common method called by models, just like the compute_output_stat.