Add caching lazy MinkowskiSum

[schillic]

In Reachability.jl we can build lazy sums in a loop. In each loop iteration we query the support vector of the sum and add one more element. It would be more efficient if the sum stored the support vector result for most summands and only computed the support vector of the remaining set(s).

Example:

sum = CSum([X1, X2, X3])
σ(d, sum)
# in the background, `sum` computes `r1 = σ(d, X1) + σ(d, X2) + σ(d, X3)`
# and stores ``[X1, X2, X3] => [(d, r1)]`
sum = sum + X4
# in the background, `sum` stores `[X1, X2, X3] => [(d, r1)], [X4] => []`
σ(d, sum)
# in the background, `sum` computes `r2 = r1 + σ(d, X4)`
# and stores `[X1, X2, X3, X4] => [(d, r2)]`

It could be helpful to have #269 first.

[mforets]

👍 i think this approach can also useful for other set representations. as a design strategy, in a future revision we could use composition like:

struct CachedLazySet{N, ST<:LazySet{N}, CT} <: LazySet
    X::ST
    cache::CT
end