Optimize data movement for CudaRosenbrocksolver and compute minus Jacobian matrix on GPU

[sjsprecious]

This PR moves the data movement of array of indices for the diagonal elements of Jacobian matrix to the constructor/destructor of CudaRosenbrocksolver class.

In addition:

• currently the operation of negating Jacobian matrix is done on the CPU (https://github.com/NCAR/micm/blob/main/include/micm/solver/cuda_rosenbrock.hpp#L94-L95) for the AlphaMinusJacobian function. This PR moves this operation to the AddJacobianTermsKernel function so the CUDA kernel now returns -J ranther than J matrix. The CPU version is updated as well. The current implementation works only when the initial Jacobian matrix is set to zero for all the elements.
• the AlphaMinusJacobianKernel function is revised to achieve more parallelism on GPU.

All the 41 tests passed on Derecho's A100 GPU with nvhpc/23.7.

Fix #391

[codecov-commenter]

Codecov Report

All modified and coverable lines are covered by tests :white_check_mark:

Project coverage is 94.66%. Comparing base (570f6ec) to head (50f50ae).

Additional details and impacted files

```diff @@ Coverage Diff @@ ## main #436 +/- ## ======================================= Coverage 94.66% 94.66% ======================================= Files 38 38 Lines 3057 3057 ======================================= Hits 2894 2894 Misses 163 163 ```

:umbrella: View full report in Codecov by Sentry.
:loudspeaker: Have feedback on the report? Share it here.