Reduce memory of the parameters class object

[jdebacker]

For a typical calibration, the parameters class object is not huge (~20MB), but if one were doing large scale parallelization (or running on a machine with low memory), this could cause issues.

I think we can reduce the footprint of this object without loosing much. A couple of ideas:

1. Use the more efficient data types (e.g. retirement_age can be cast as an integer, more floating point values would be fine with 32 bit (or even 16 bit).
2. Most parameters become constant long before the steady state: is there a way to reduce the size of arrays knowing that after some point in the T dimension, things stop changing?

[jdebacker]

Related: Example of custom data type in ParamTools