Refactor: Centralize and Extend update_model_kwargs_for_generation Logic

[w-yyh]

Centralized Highly Common Logic: The following logic, which was duplicated across multiple model subclasses, has been moved into the universal update_model_kwargs_for_generation function:

position_ids update: Logic to increment and append the next position ID for self-attention models.

4D attention_mask support: Extended attention_mask handling to correctly update 4-dimensional masks during generation (in addition to the existing 2D support).

Delete the redundant code in models.The changes were applied only to pure text models, with the exception of Llama.

[paddle-bot]

Thanks for your contribution!

[codecov-commenter]

Codecov Report

❌ Patch coverage is 66.66667% with 3 lines in your changes missing coverage. Please review.
⚠️ Please upload report for BASE (develop@9594389). Learn more about missing BASE report.

Files with missing lines	Patch %	Lines
paddleformers/generation/utils.py	62.50%	3 Missing ⚠️

❌ Your patch status has failed because the patch coverage (66.66%) is below the target coverage (80.00%). You can increase the patch coverage or adjust the target coverage.

Additional details and impacted files

@@            Coverage Diff             @@
##             develop    #2972   +/-   ##
==========================================
  Coverage           ?   31.38%           
==========================================
  Files              ?      420           
  Lines              ?    68405           
  Branches           ?        0           
==========================================
  Hits               ?    21469           
  Misses             ?    46936           
  Partials           ?        0           

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

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.