Filter small coefficients and handle numerical precision

[DillonJ]

Sometimes models can have small coefficients that may as well be zero but aren't quite. An almost-zero coefficient might have negligible impact on a model solution but might impact run time significantly. There is value in eliminating these as solvers make use of sparse matrix methods extensively.

Additionally, in printed outputs from Julia, I am constrantly seeing things like this: 0.050000000000000003. The value specified was 0.05 but somewhere, numerical precision issues are resulting in this value being slightly imprecise. We have seen this happen with zeroes/non-zeroes too... a value that should be zero somehow ends up being very small instead.

Here is another related issue: spine-tools/SpineOpt.jl#585

In summary, I think we should filter and/or round coefficients before they are sent to the solver to make maximum use of zeroes.

Perhaps two arguments could be specified in the "using_spine_db" function. One would be the lowest absolute value and any float (whether scalar or in a timeseries or map) with an absolute value of less, is treated as zero. The second argument could be a global precision argument that rounds all floats to that level of precision.

Thoughts @manuelma? I would at least like to test this as I suspect it could have a big impact on some models.