Fix unit_diagonal and non-contiguous reads in MLX linalg

[jessegrabowski]

MLX's solve_triangular has no unit_diagonal, and the dispatch dropped the flag rather than emulating it, so it solved against A's real diagonal. lu_solve hit this too, since it hands the packed LU factor straight in and relies on the flag to ignore U's diagonal.

Second, unrelated: mx.array keeps a non-C-contiguous numpy layout and MLX's elementwise kernels misread the buffer, so mlx_typify normalizes to C order. That made lu_solve's gradient wrong, and reverting it segfaults rather than failing on the numbers, so the read is going out of bounds.

Closes #2384