Optimization for moving least squares

[mrlag31]

This PR modifies the SVD pseudo-inverse and the coefficients computation, its main goal was to reduce the amount of global memory used, use scratch memory and compound the kernels.

[aprokop]

@mrlag31 Can you please explain in more details what the difference is?

[mrlag31]

• The previous SVD implementation used two extra square matrices (one for the diagonalisation, one for the unit matrices). Now, it uses a vector and a square matrix, the diagonalisation is done on the input matrix.
• The coefficients computation is now batched per-thread using a team policy. All of the previous kernels were compounded into a single one and all of the temporary matrices are now placed in shared memory.

[dalg24]

Drive-by comments

[aprokop]

I'm not sure about renaming in symmetricPseudoInverseSVDKernel. I don't think the new names make it any clearer, and would prefer the old names.

[mrlag31]

I'm not sure about renaming in symmetricPseudoInverseSVDKernel. I don't think the new names make it any clearer, and would prefer the old names.

I felt like the old ones were harder to understand. The new ones directly explains, by their name, what they are in regards to the formula, whereas the old ones were a bit too abstract.

[aprokop]

I'm ok with the current state.