Use create_bidirectional_mask for backend-agnostic attention mask handling

fix: align _init_weights with Qwen2Moe using nn.init API

fix: call super()._init_weights() to match Qwen2Moe convention for transformers v5

fix: align RotaryEmbedding with Qwen2Moe pattern for transformers compat

Update README.md

Update README.md

Update modeling_llada2_moe.py

Update README.md

Create modeling_llada2_moe.py

Update configuration_llada2_moe.py

Create README.md

Add files using upload-large-folder tool

initial commit