Tensor-Compressed and Fully-Quantized Training of Neural PDE Solvers