fix(modules): guard HF modeling patches behind torch availability

[Gpgabriel25]

Problem

Five model configuration modules unconditionally call
AutoModel.register / AutoConfig.register at import time, requiring
PyTorch to be installed even in pure-JAX environments. In any torch-free
deployment (TPU pods, CI without torch, EasyDeL-as-library):

ImportError: modeling_glm4v_moe requires the PyTorch library but it was
not found in your environment.

All five modules fail immediately on import:

• glm4v_moe_configuration
• llama4_configuration
• phimoe_configuration
• qwen3_5_moe_configuration
• qwen3_next_configuration

Fix

Wrap the registration calls with HuggingFace's standard
if is_torch_available(): guard.

Evidence

Import probe, same torch-free venv, main vs this branch:

module	main	this PR
glm4v_moe_configuration	ImportError	✅ OK
llama4_configuration	ImportError	✅ OK
phimoe_configuration	ImportError	✅ OK
qwen3_5_moe_configuration	ImportError	✅ OK
qwen3_next_configuration	ImportError	✅ OK