train with `capacity_mwp` instead of `installedcapacity_mwp`

[peterdudfield]

Detailed Description

Normalize gsp data by capacity_mwp rather than installed.

Would be good to do this as an exepreince to see if this improves the results or not

Context

• capacity_mwp is a better estimate of the current live amount of solar data there is

Possible Implementation

change in inference and training