Avoid the local copy of Jacobian matrix when doing LuDecompose

[sjsprecious]

According to the Nsight Systems profiling results, there is a huge gap between the SubtractJacobianTerms and AlphaMinusJacobian kernels. The root cause is that the line auto jacobian = state.jacobian_ calls the default copy assignment operator, and it will copy all the data member on the host (https://github.com/NCAR/micm/blob/main/include/micm/util/sparse_matrix.hpp#L52-L56). This is unnecessary if we just want to copy the data from device to device

Acceptance Criteria

• Use state.jacobian_ directly as the function argument.
• Call the SubtractJacobianTerms function to re-construct the Jacobian matrix if the matrix is singular and substepping is needed. The computational time is much shorter than making a copy of the Jacobian matrix based on the profiling result.