Fixes to thickness regularization

[gillian-smith]

Here I propose a few fixes/changes/additions to the thickness regularization in the data_assimilation module.

regu_thk_v1()

• when setting processes.data_assimilation.regularization.smooth_anisotropy_factor less than 1.0, ave4() needs to be applied to make dims of state.flowdirx, state.flowdiry match dims of dbdx, dbdy

regu_thk_v2()

• The 2nd derivative term has a term missing, which is required to make it rotationally invariant/independent of choice of $x$ and $y$ axes: see https://hackmd.io/@gillian-smith/r1luUbVh-g for derivation (many thanks to James Maddison for spotting this)

• The 1st derivative term is isotropic even if processes.data_assimilation.regularization.smooth_anisotropy_factor < 1.0 - I have changed this to match behaviour of regu_thk_v1()

• For the 1st derivatives, the finite difference kernels kx and ky decouple odd- and even-indexed cells. This induces chessboard modes when the 2nd derivative term is inactive (processes.data_assimilation.regularization.thk_2nd_der=0). I propose reverting to the discretization used for regu_thk_v1(), or at least give users the option, with new parameter processes.data_assimilation.regularization.thk_1st_der_disc_version which can be "new" or "old".

• Currently the kernel kxy also decouples odd- and even-numbered cells, but since bxy is always combined with other terms, this appears to not produce any visible numerical artefacts. I have left this as-is, but if we would prefer to change it, perhaps kxy = tf.constant([[ 1., -1., 0.],[ -1., 2., -1.],[0., -1., 1.]], tf.float32) / (2.0*dx*dx) is an option? (see https://en.wikipedia.org/wiki/Finite_difference#Multivariate_finite_differences)

• suggest naming the coefficient variables alpha_2 and alpha_1 instead of alpha and beta, just because $\beta$ represents the anisotropy factor in https://doi.org/10.1017/jog.2022.41 and https://doi.org/10.5194/egusphere-2026-788, and it would be less confusing to keep this notation

After these fixes,
processes.data_assimilation.regularization.thk_version=2 processes.data_assimilation.regularization.abl_acc_balance=1 processes.data_assimilation.regularization.thk_2nd_der=0 processes.data_assimilation.regularization.thk_1st_der_disc_version=old

has effectively the same behaviour as
processes.data_assimilation.regularization.thk_version=1,

and the parameters processes.data_assimilation.regularization.thk_1st_der and processes.data_assimilation.regularization.thk are equivalent (up to a constant factor).

I think the regu_thk_v1() and regu_thk_v2() can be unified into one function, but perhaps later in another pull request.