Fix global PyTorch initializer leak in build_vae_var

[primorLee]

Summary

• scope the temporary PyTorch reset_parameters overrides to VAR/VQVAE construction
• restore every original initializer in a finally block on both success and failure
• add CPU regression coverage for normal construction and constructor exceptions

Problem

build_vae_var() disables built-in initialization by replacing reset_parameters on eight global torch.nn classes. Those class methods were never restored. After one VAR build, unrelated modules created anywhere in the same Python process could therefore keep uninitialized weights.

A minimal CPU reproduction on current main prints linear_reset_restored=False after build_vae_var() returns.

Fix

Save the original class methods, keep the existing no-initialization fast path while VQVAE and VAR are constructed, and restore the methods in finally. The existing explicit var_wo_ddp.init_weights(...) remains unchanged.

Validation

• real reduced CPU VAR/VQVAE construction verifies all eight PyTorch initializers are restored
• injected VQVAE construction failure verifies restoration on the exception path
• python -m pytest tests/test_model_initialization.py -q (2 passed, 16 subtests)
• python -m compileall -q models/__init__.py tests/test_model_initialization.py
• repeated both regression cases in a clean Linux Python 3.10 / CPU Torch 2.5.1 container

Fixes #159