Fix multithreading reproducibility

[lrobion]

Closes #19. Multithreaded simulations are now bitwise reproducible, adds CI checking bitwise reproducibility between two runs (1 and 4 threads) to ensure this does not happen again.

Multithreaded simulations even using a fixed random seed were not reproducible. This was for three reasons:

1. Multithreaded access to the global RNG state in updating the temperature perturbation. Even if the RNG state is consistent, the threads access it in a non-deterministic way making the temperature perturbation different even run.
2. Plain race condition in a conversion factor in LAGRIDPlumeModel.cpp and two in Aerosol.cpp (including the moment calculation)
3. Parallel reductions in Aerosol.cpp that due to floating point non-associativity lead to different results. This floating point noise is magnified because the shear applied to the contrail depends on finding the maximum index of xOD with depends on multiple non-deterministic reductions. Eventually the numerical noise is sufficient to shift the index of the max xOD which changes the shear which then leads to a different simulation grid.

This addresses all 3 issues but comes at a signficant performance drop (40% slower simulations on 4h of the ISSL 140 test) due to the non-parallel reductions (see new comments). This performance drop can be reduced further if we don't repeat these (now slower) reductions multiple times per timestep (see #97).

I think once we address the repeated calculations, we can decide if the performance drop is still significant enough that want to look into this more. Further work on the performance of this should probably happen after #92 is merged as it also deals with performance in Aerosol.cpp.

[lrobion]

Rebased on the #92, added back safe parallelism... The slowdown is negligible on my machine (issl_140 example for 4h, mean of 3 runs):

Threads	Main (s)	New (s)	Change
1	195.8	198.6	+1.4%
2	121.4	120.1	-1.1%
4	76.9	76.7	-0.2%
8	74.6	74.7	+0.1%

All results are bitwise identical with main running on 1 CPU. For the moment calculations I changed the loop order to make parallelism safe where each thread computes a row of moments in the 2D grid so they do not overlap but still keep continuous data access patterns in the innermost loop