2.47.00: Bug fix for complex forms

Earlier this week, an API user alerted us to an issue with the stacked_form
 endpoint where there were (pun intended) a few stacked/compounding issues. This release addresses these issues.
Diagnosis of the issue
A user approached us with two building examples which included multiple towers over a podium. The structural emissions estimates for that form were lower than expected, and the difference between the two options was wider than anticipated. A few of the issues we identified:
A high-rise mass timber building, which has few built or designed precedents. This stretched our background dataset to the very the edge.
A formally complex stack, which the model separated out into a number of smaller vertically-bearing portion of the building (C.Scale affectionately refers to these portions of a building as chunks).
The combination of multiple use types within a single chunk, which required partitioning of the chunk into multiple model runs to capture patterns by use type.
In reviewing the error, we identified an issue in how we were handling complex building forms with multiple use types in a single stack of levels. By performing a set of operations on each chunk individually, the algorithm which interpreted the building form was not tracking how these forms should be re-combined into a whole. This error was identified because the area-weighted SMQi for all chunks was not equal to the SMQi exposed in the API response.
In many cases, this error is small (within a few percent of the total) but, in this case, the three points above led to a more significant deviation. The combination of all these issues led to an underprediction of emissions and an unexpected scale of difference in prediction between a simpler and a more complex building form.
C.Scale's Response
To fix this issue, C.Scale made two changes -
Previously, we removed mass timber projects with very high steel quantities from our background data, leading to underpredictions for buildings of more than a few stories but more reliable predictions for low-rise buildings. We have reintroduced some mass timber buildings with higher steel quantities into the dataset to allow more reasonable extension of the data to mid-rise buildings. However, we note that this remains a data-scarce area of the mode.
The stacking algorithm has been rewritten to enforce that results for from each chunk of the building retain correspondence to the building as a whole.
Regression
After these fixes have been implemented, we see a change of -5% to +18% difference in emissions predictions for mass timber projects, and larger differences for complex stacked forms. These changes represent improvements to the model and bug fixes. 