[FEAT] add code of PhysicsRegression model

[lkyu-ly]

This example integrates the Paddle version of PhysicsRegression into PaddleCFD for the paper A Neural Symbolic Model for Space Physics.

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

Loss comparison of a training test on a rather small config than default:

froward diff:

logits_shape: (7, 10570)
torch_dtype: float32
paddle_dtype: float32
mean_abs_error: 5.4308799236e-07
max_abs_error: 5.7220458984e-06
mean_rel_error: 1.6366435602e-06
max_rel_error: 3.4902596381e-03

[lkyu-ly]

To switch to _SimpleIterator for deterministic cross-framework loss alignment:

1. Comment out the return paddle.io.DataLoader(...) block above.
2. Uncomment the class _SimpleIterator and the final return _SimpleIterator(...) below.
3. If aligning against a PyTorch baseline, the PyTorch side MUST apply the same change
   (replace its DataLoader with an equivalent _SimpleIterator) — otherwise batch order and
   RNG state will diverge between the two frameworks, making per-step loss comparison invalid.

Loss alignment results (first epoch, sampled every 50 steps, PyTorch vs Paddle):
Shape: (10,) | dtype: float32 vs float32
max_abs_diff: 0.007700 mean_abs_diff: 0.003450 rmse: 0.004042
max_rel_diff: 0.043800 mean_rel_diff: 0.019583
diff range: [-0.007700, 0.003900]