[Qwen3_VL] use precomputed visual scatter indices in deepstack process

[huangjiyi]

Summary

• Use precomputed visual_update_indices and visual_gather_indices (from PaddleFleet GPTEmbedding, see [VL] precompute visual scatter indices in GPTEmbedding PaddleFleet#708) in _deepstack_process to avoid redundant per-layer index computation.
• Retains full fallback path for backward compatibility when precomputed indices are not provided.

Test plan

• Verify Qwen3_VL pretrain/SFT training produces identical results
• Confirm no regression in training loss/performance
• Test with and without PaddleFleet#708 to validate fallback path

[paddle-bot]

Thanks for your contribution!

[codecov-commenter]

Codecov Report

❌ Patch coverage is 0% with 43 lines in your changes missing coverage. Please review.
⚠️ Please upload report for BASE (develop@f012f4f). Learn more about missing BASE report.

Files with missing lines	Patch %	Lines
...dleformers/transformers/qwen3_vl/modeling_fleet.py	0.00%	43 Missing ⚠️

❌ Your patch status has failed because the patch coverage (0.00%) is below the target coverage (75.00%). You can increase the patch coverage or adjust the target coverage.

Additional details and impacted files

@@            Coverage Diff             @@
##             develop    #4189   +/-   ##
==========================================
  Coverage           ?   33.89%           
==========================================
  Files              ?      462           
  Lines              ?    87499           
  Branches           ?        0           
==========================================
  Hits               ?    29661           
  Misses             ?    57838           
  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.