Gemma 4 models cannot use

version: `ddf1b879dc3a1760cbcb3f3c4a7c6467850cec4a`
env: Ubuntu 24.04; CUDA 12.8

**Problem 1: Gemma 4 text checkpoint does not load**
The gemma4 example looks up `model.embed_tokens.weight`, but official Gemma 4 checkpoints store the language-model weights like `model.language_model.*.` On google/gemma-4-E4B-it that fails with cannot find tensor `model.embed_tokens.weight`

**Problem 2: Gemma 4 E2B/E4B model does not work as it output random words**

Start by running the gemma4 example with the google/gemma-4-E4B-it checkpoint on the reported Ubuntu and CUDA environment, then inspect the model tensor lookup and generation behavior. Done means the official Gemma 4 E2B/E4B checkpoints load successfully and produce meaningful output rather than failing on model.embed_tokens.weight or generating random words.

Written by the indexing model from the issue text.