Implement `collision_probability` for LSHFunction subtypes

[kernelmethod]

SimHash, MinHash, and LpHash currently all have implementations of single_hash_collision_probability (MIPSHash and SignALSH are currently missing them, see issue #4). There should also be a function collision_probability(::LSHFunction, sim) that computes the overall probability that the total hash (i.e. the vector of individual hashes) generated by an LSHFunction is equal for two inputs with similarity sim.

Note that if single_hash_collision_probability is already implemented then this is relatively easy to implement using

collision_probability(hashfn::LSHFunction, sim::Real) =
    single_hash_collision_probability(hashfn, sim)^nhashes(hashfn)

[kernelmethod]

This is fixed by 41e6c45. I've added two different functions to the LSHFunction API:

collision_probability(hashfn::LSHFunction, sim;
                      n_hashes::Union{Symbol,Integer} = :auto)

This implementation of collision_probability computes the probability of n_hashes simultaneous collisions between a pair of inputs with similarity sim. If n_hashes is left as the default :auto, then we'll compute the probability for n_hashes(hashfn) simultaneous hash collisions (using the n_hashes function exported by the LSHFunction API).

collision_probability(hashfn::LSHFunction, x, y; kws...)

This is for convenience as it's just equivalent to collision_probability(hashfn, similarity(hashfn)(x,y); kws...). E.g. we have

julia> hashfn = MinHash(10);

julia> A = Set(["a", "b", "c"]);

julia> B = Set(["b", "c", "d"]);

julia> collision_probability(hashfn, A, B) ==
       collision_probability(hashfn, similarity(hashfn)(A,B)) ==
       collision_probability(hashfn, jaccard(A,B))
true

For the LSHFunction internal API, LSHFunction subtypes must implement

single_hash_collision_probability(hashfn, sim)

which is the probability that a pair of inputs with similarity sim experiences a hash collision for just one hash function.