Fix _init_weights and RotaryEmbedding for transformers v5.x compatibility
Fix _init_weights and RotaryEmbedding initialization (for transformers 5.x)
_init_weights was using .data.normal_() directly on tensors, which bypasses the _is_hf_initialized guard in transformers v5.x. Since v5.x loads on meta device first then calls initialize_weights() post-checkpoint, this was silently re-randomizing every Linear and Embedding after from_pretrained. Model loads fine, outputs garbage. Switched to torch.nn.init.normal_() / zeros_() so the guard works.
Also, RotaryEmbedding.__init__ KeyErrors on "default" rope type - ROPE_INIT_FUNCTIONS just doesn't have that key, and Ring-mini-2.0 has rope_scaling=None so it always hits this path. Handled default inline. While at it, forced float32 for the inv_freq computation because rope_theta=600k overflows bf16 trivially.