Wrong device for residual-based weights in RBAPINN

[GiovanniCanali]

Describe the bug
When training with RBAPINN, if the selected accelerator is not cpu (e.g., "cuda"), the residual-based weight buffers are not properly moved to the correct device. As a result, a device mismatch error occurs during training, preventing execution from continuing.

To Reproduce

from pina.problem.zoo import Poisson2DSquareProblem as Poisson
from pina.model import FeedForward
from pina.solver import RBAPINN
from pina import Trainer

# Define the problem
problem = Poisson()
problem.discretise_domain(n=50, mode="grid")

# Define the model and the solver
model = FeedForward(2, 1, 64, 2)
solver = RBAPINN(problem=problem, model=model)

# Initialize the trainer and begin training
trainer = Trainer(solver=solver, max_epochs=100, accelerator="cuda", devices=1)
trainer.train()

Expected behavior
Training should proceed without device mismatch errors, and all internal buffers should be moved automatically to the specified accelerator prior to training.