Support GLPK use in backfill corrections container

[nmdefries]

At the moment, backfill corrections can't run in the docker container without a Gurobi license, provided via params.json. If license info isn't provided or if the user requests it, we want the process to be able to use GLPK instead. Which optimizer to use should be available as a setting in params.

[nmdefries]

Using GLPK to fit models also isn't much/any slower than gurobi, so if licensing becomes a problem in the future, it is not a big deal to switch. The intuition here is that gurobi is (apparently) optimized to fit "big"/"complex" models quickly but that those optimizations don't necessarily apply well to all or perhaps some types of smaller problems.

I've timed a couple local runs on smaller datasets to look at speedup of gurobi over GLPK, so nothing too formal or in-depth. There's a 0-3 s increase in runtime using GLPK over gurobi. Changing the solver could have a bigger impact when using more data/longer lookbacks, like we'd do in production.

[nmdefries]

Most of this is already set up, that is we have a params$lp_solver setting that is properly passed around to functions as needed. However, we don't gracefully handle cases when gurobi is requested but can't be used (e.g. license info is invalid). In those cases, we get Training failed for cache/20200221_20200122_changehc_covid_state_lambda0.1_count_ak_lag1_tau0.01.model, for example, and cascading errors like Error in integer(n) : invalid 'length' argument.