feat: inner products for mixed method

[chauj96]

This PR adds the implementation of computeM, which computes the inner product matrix $M$ used in the MFD mixed method formulation.

1. The function has been implemented for the following inner product types:

• TPFAInnerProduct
• BdVLMInnerProduct
• QuasiTPFAInnerProduct
• SimpleInnerProduct
  Each class now includes a computeM routine that construct the inner product matrix $M$ required for the mixed formulation.

2. Testing
   Unit tests were added in testMimeticInnerProducts to verify the correctness of the newly implemented computeM functions. All tests currently pass.

3. The centroid computation has been updated to evaluate the exact volume of a polyhedron with planar polygonal facets. This property is verified by a unit test. This induces a rebaseline.

+) Minor variable naming updates were applied in SimpleInnerProduct to keep the geometry nomenclature consistent with the other inner product implemenations.

[OmarDuran]

Thank you @chauj96, nice work.

[bd713]

Let's update the PR description to clearly explain that the change in centroid computation is the reason a rebaseline is needed.

[jafranc]

Great work on rewriting as inner products rather than transmissivities.

Although I am unclear on why it is not called in compute replacing the old way and just duplicated from it up to the transposition.

[OmarDuran]

Great work on rewriting as inner products rather than transmissivities.

Although I am unclear on why it is not called in compute replacing the old way and just duplicated from it up to the transposition.

Thanks! The reason computeM doesn't replace the body of compute is that they build different matrices: compute assembles the inverse inner product W = (N K Nᵀ + t·P_C diag(N K Nᵀ) P_C)/V for the hybrid (flux-elimination) scheme, while computeM assembles the primal M = C K⁻¹ Cᵀ/V + (V/t)·P_N diag(N K Nᵀ)⁻¹ P_N for the mixed saddle-point formulation, where fluxes stay unknowns. The dual structure of M↔W (C↔N, K↔K⁻¹, t↔1/t) exists by consistency construction, but the stabilizations differ, so in general M ≠ W⁻¹, if we want to call one from the other would need a dense per-element inversion in-kernel. The actual shared part is now factored into computeCellToFaceGeometry (C, N, face areas) and computeConsistencyTerm (C K⁻¹ Cᵀ/V).

[jafranc]

Great refactor 👍