[WIP] Use Tullio for pairwise distances

[theogf]

Summary We have a long-time problem for binary operations like DotProduct not satisfying the requirements of the Distances.jl framework (not a proper metric). Additionally, Distances.jl is very incompatible with GPU operations (see JuliaStats/Distances.jl#143 and JuliaStats/Distances.jl#137). Using Tullio.jl should solve both these problems. Some quick benchmarks shows that Tullio is both faster and more GPU-able than Distances.jl

There is a longer discussion about this PR in #380 This should also close #98 and replace #194

Proposed changes

• Tullio is used for computing all pairwise and colwise
• The only thing used from Distances.jl are their types (we stop using Distances.pairwise).
• Adding special implementations of pairwise for ColVecs and RowVecs when possible to improve speed (and GPU compatibility)
• Create a AbstractBinaryOp abstract type for objects like DotProduct and Delta and combine them with Distances using BinaryOp = Union{AbstractBinaryOp,Distances.PreMetric}.

What alternatives have you considered? Dropping Distances.jl operations anyway but without Tullio but Tullio shows it's faster.

Breaking changes

• It's technically not breaking since the API does not change, only the internal computations change

[theogf]

Ah... Ups I think I did not branch from master....

[theogf]

Blessed be git cherry-pick