feat: support MOSS split QKV weights and batch host loading

[leejet]

Summary

• Map MOSS split Q/K/V weights to fused QKV tensor parts.
• Batch tensor parts stored in host memory with ordinary weights, allowing parallel loading and conversion.

Related Issue / Discussion

Fix #2158.

Additional Information

N/A

Checklist

• I have read and confirmed this PR follows the contribution guidelines.

Summary by CodeRabbit

• Bug Fixes
  • Improved loading of compatible tensors from shared sources, including tensors assembled from multiple parts. Loading now reports failures when setup, size checks, or data loading fail.
• New Features
  • Added support for mapping Moss Image attention projection weights to the combined QKV naming format.

[coderabbitai]

Note

Currently processing new changes in this PR. This may take a few minutes, please wait...

⚙️ Run configuration

• Configuration used: Repository UI
• Review profile: CHILL
• Plan: Advanced
• Run ID: 1628773d-1611-4221-9c6e-c96d05f0a062

📥 Commits

Reviewing files that changed from the base of the PR and between ab8d5ea and 81deae8.

📒 Files selected for processing: 2

• src/model_manager.cpp
• src/name_conversion.cpp

+-----------------------------------------+
| Time zones: the final boss of software. |
+-----------------------------------------+
  \
   \     /)/)
        ( '')
        /  |
      o(___)

✨ Finishing Touches

📝 Generate docstrings

• Commit to this branch
• Create a new PR

🧪 Generate unit tests (beta)

• Commit to this branch
• Create a new PR

• Autofix · Keep fixing CodeRabbit findings and required CI, and resolving merge conflicts

_{Comment @coderabbitai help to get the list of available commands.}