fix(qwen3_5): skip GDN input-projection fusion on mixed projection dt…

[potoior]

…ypes

Anti-compressed NVFP4 checkpoints can quantize only some GDN input projections (e.g. in_proj_qkv/in_proj_z -> NVFP4 FP8 weights) while leaving others (in_proj_b/in_proj_a) FP16. fuse_gdn_input_projections checked only the first projection's quant type and then concat weights of all four projections, so these checkpoints crashed with 'Promotion for Float8 Types is not supported, attempted to promote Float8_e4m3fn and Half'.

Verify all four projections share the same weight dtype before fusing; when they differ, log a warning and keep the unfused path (which is already supported in forward()). Pure FP16 / pure NVFP4 models still fuse.

Verified: tensorrt-edgellm-export Qwen3.8-27B-NVFP4 --skip-audio completes, with 'GDN fusion skipped' warnings logged per layer.

What does this PR do?

Type of change: ?

Overview: ?

Usage

# Add a code snippet demonstrating how to use this

🚀 Pull Request Checklist

Thank you for contributing to TensorRT Edge-LLM! Before we review your pull request, please make sure the following items are complete.
Please also refer to Contributor guidelines for general guidelines.

✅ Pre-commit Checks

• I have installed pre-commit by running pip install pre-commit.
• I have installed the hooks with pre-commit install.
• I have run the hooks manually with pre-commit run --all-files and fixed any reported issues.

🧪 Tests

• Tests have been added or updated as needed.
• All tests are passing.

📄 Documentation

• Updated any necessary documentation

⚙️ Compatibility

• The change is backward compatible

Additional Information