[FEAT] add Multiple Physics Pretraining (MPP) model

[lkyu-ly]

• integrate AViT (Axial Vision Transformer) with MPP normalization into ppcfd.models
• port all model architecture, custom optimizers (Adan, DAdapt), and data loaders from PaddlePaddle migration
• add training, inference, and weight conversion scripts under examples/
• support model variants Ti/S/B/L with PDEBench multi-physics datasets (SWE, NS, diffusion-reaction)

This example is able to run on both Nvidia GPU and Custom Device(using iluvatar for test). CINN is also supported.

forward precision:

max_abs_error: 7.8231096268e-07
mean_abs_error: 9.9473159783e-08
rmse: 1.3069938825e-07
max_rel_error: 4.3364424808e-03
mean_rel_error: 6.9402936241e-06