Validate dt against CFL stability bound in diffuse() (#1281)

[brendancol]

Summary

• diffuse() validated user-supplied dt only as > 0. The explicit forward-Euler 5-point Laplacian is unconditionally unstable above the CFL bound dt = 0.25 * dx**2 / max(alpha), and steps is accepted up to 100,000, so an oversized dt silently produced compounding Inf/NaN with no error.
• The dt=None branch already used this bound. This PR hoists it into cfl_max and raises ValueError when the user-supplied dt is larger.
• The check runs in the public entrypoint, so it covers all four backends (numpy, cupy, dask+numpy, dask+cupy).

Test plan

• dt above the bound raises ValueError
• dt at the bound is accepted
• dt=None still works
• CFL check uses max(alpha) for spatially varying diffusivity
• all 19 tests in test_diffusion.py pass

Closes #1281