Fix lumina2 pad token shape mismatch for some GGUF conversions

[vaclavmuller]

This PR fixes a shape mismatch when loading some lumina2 / NextDiT GGUF models
(e.g. Z-Image Turbo GGUF builds).

Some GGUF conversions store x_pad_token and cap_pad_token as 1D vectors
([D]) instead of the expected 2D shape ([1, D]), which causes
load_state_dict to fail.

The loader now:

• ensures a robust fallback shape when orig_shape metadata is missing
• reshapes lumina2 pad tokens to (1, D) when needed

Tested with:
https://huggingface.co/leejet/Z-Image-Turbo-GGUF

Addresses #379