ENH add "split" discretization

Following:
- [Chambolle, Antonin, and Thomas Pock. "A remark on accelerated block coordinate descent for computing the proximity operators of a sum of convex functions." The SMAI Journal of computational mathematics 1 (2015): 29-54.](http://www.numdam.org/article/SMAI-JCM_2015__1__29_0.pdf)
- [Chambolle, Antonin, Pauline Tan, and Samuel Vaiter. "Accelerated alternating descent methods for Dykstra-like problems." Journal of Mathematical Imaging and Vision 59 (2017): 481-497.](https://hal.science/hal-01346532/file/AlterDescent2.pdf)

This is an alternative discretisation of the continuous 2D TV friendly to GPU computation.

No file, test, or entry point is named in the issue. Start by locating the existing 2D TV discretization in the Python repository, then read the two cited papers to define the GPU-friendly split formulation; done means the alternative discretization is implemented and its behavior is validated against the benchmark's existing approach.

Written by the indexing model from the issue text.