Make model training fully deterministic

[jeancochrane]

Our investigation in https://github.com/ccao-data/enterprise-intelligence/pull/258 revealed that two model training runs with identical hyperpameters can produce slightly different tree structures, with enough variation to lead to ~80 cards with significant prediction differences.

We should already be using lightgbm in a deterministic fashion, so there are two main possibilities I see:

• Our determinism configuration is not quite correct, and needs to be tweaked (see deterministic and related flags don't guarantee same result on different machines microsoft/LightGBM#6683)
• There's a bug in lightgbm core

I think the trickiest part of this issue will be generating a reproducible example that we can use to test and confirm that our fix worked.