Refactor Model and UniversalFallback

[blegat]

Closes #1234

[odow]

So this is actually slower than before, when run on the JuMP examples (The stress_precompile function here: jump-dev/JuMP.jl#2484).

Before:

 78.339035 seconds (204.57 M allocations: 11.833 GiB, 5.22% gc time, 6.89% compilation time)

After:

88.580525 seconds (220.25 M allocations: 12.709 GiB, 5.61% gc time, 14.41% compilation time)

However, it is more readable. I'd be in favor of merging so we can make other changes in future PRs. There are likely some performance things we can tweak.

[blegat]

Thanks for doing a benchmark. I think there is an issue with Fix1 and CleverDict as we have Fix1{DataType, ...} which does not play well with inference.