Let the element type and the endpoint types be different

The rough conclusion that we reached in #117 is that the endpoint types should not necessarily be the same as the eltype.

I decided to implement this in a PR that I will submit shortly, but am filing this issue for conceptual discussion. The key lines that summarize the new implementation are:
```julia
struct Interval{L,R,T,TL,TR} <: TypedEndpointsInterval{L,R,T}
left::TL
right::TR
end
```

There are some important choices I made with my implementation:
1. I pushed the endpoint types as far down in the type hierarchy as possible, i.e. only the concrete type `Interval` has endpoint types. There are no abstract types with endpoint type parameters. The type of the endpoint can be accessed with e.g. `typeof(leftendpoint(i))`. This seemed to be the simplest and lowest-risk option, and we can change it later if necessary.
2. One choice is what the default eltype should be **if the user doesn't specify it explicitly**. I originally tried to keep the `eltype` as close to the endpoint types except for `Integers` (e.g. `eltype(1..3//2)` would be `Rational`, `eltype(1..2)` would be `Float64`), but this ended up feeling too special-case-y. I decided a better option would be to call `float` to determine the `eltype` from all endpoint types `<:Number` in the `Interval` constructor (i.e. `eltype(1..3//2)` is `Float64`). This ended up feeling very nice and consistent and works well with other packages like Unitful.
3. I also added a docstring defining the meaning of `eltype` for a `Domain`. It is "the type that best represents elements of the domain according to the criteria chosen by the programmer who created the domain." I think it is best to give users flexibility since it is hard to predict everything `Domain`s will be used for.

What does everyone think?