Add live training monitor

[fvalle1]

What's missing

DeepXDE has no way to watch live a PINN train. The closest existing tools are:

• MovieDumper — accumulates frames but only writes the GIF in
  on_train_end, after training has already finished.
• dde.saveplot() — plots the solution and loss history, but only once,
  after training completes.

Neither gives feedback while the model is still training. Other PINN
libraries (e.g. neurodiffeq's Monitor1D/Monitor2D)
support this: a periodically-refreshed plot of the current
prediction and loss history during training, which is very useful for
spotting divergence, bad loss weighting, or stalled optimization early instead of after a long run.

This feature can be helpful for issues (#2014, #1116).

What this PR adds

• dde.callbacks.TrainingMonitor: a new Callback that, every period
  epochs, redraws subplots — the current predicted solution
  (optionally against a reference/analytical solution via y_reference)
  and the train/test loss history on a log scale.
• Designed to be safe in headless/CI environments: it detects when there
  is no usable interactive Matplotlib backend/display and disables itself
  gracefully.
• New example: examples/pinn_forward/ode_2nd_with_monitor.py, a damped
  harmonic oscillator solved with a PINN, using TrainingMonitor against
  the closed-form analytical solution.
• New example: examples/pinn_forward/diffusion_1d_with_monitor.py, a monitor added to an already existing example.
• New test: tests/test_callbacks.py, covering periodic triggering logic
  and headless-safety under the Agg backend.

Usage

monitor = dde.callbacks.TrainingMonitor(
    period=200,
    x_plot=x_plot,
    y_reference=analytical_solution,
)
model.train(iterations=10000, callbacks=[monitor])

Test plan

• pytest tests/test_callbacks.py passes (4/4) under MPLBACKEND=Agg
• examples/pinn_forward/ode_2nd_with_monitor.py runs end-to-end
• import deepxde unaffected across the change
• pylint deepxde/callbacks.py shows no new warnings beyond the
  file's existing baseline

Still in progress

• Update .rst files with documentation for ode example
• Test monitor against all existing examples