Remove PyCall as a test dependency by factoring out python model runs

For less fragile testing, I propose that the python models, which are currently run in tests to make comparisons with MLJLinearModels, be run separately from tests following the recent model of removing R models, #167.

For example, the code generating python model results now in test/fit/logistic-multinomial.jl is moved to a separate file test/fit/logistic-multinomial_py.jl; running that file serialises the results to a file called test/fit/logistic-multinomial.toml; the remaining DO_COMPARISON code in /fit/logistic-multinomial.jl reads in the data needed for comparisons from the toml file to do the comparisons.

This follows the pattern of the files in test/fit/ now called quantile.jl, quantile.r.jl and quantlile.toml (auto-generated).

Start by comparing the current Python model code in test/fit/logistic-multinomial.jl with the separated R-model pattern in test/fit/quantile.jl, test/fit/quantile.r.jl, and test/fit/quantile.toml, while reviewing issue #167. Move Python result generation to test/fit/logistic-multinomial_py.jl and have the comparison test read test/fit/logistic-multinomial.toml; done means comparisons still run without PyCall as a test dependency.