Use a more performant formula for haversine (#90)

* Use a more performant formula for haversine

```julia
using BenchmarkTools
Main> using BenchmarkTools

Main> a = rand()
0.8809105663628993

Main> @benchmark asin(sqrt($a))
BenchmarkTools.Trial:
  memory estimate:  0 bytes
  allocs estimate:  0
  --------------
  minimum time:     15.790 ns (0.00% GC)
  median time:      16.579 ns (0.00% GC)
  mean time:        17.691 ns (0.00% GC)
  maximum time:     712.929 ns (0.00% GC)
  --------------
  samples:          10000
  evals/sample:     1000

Main> @benchmark atan2(sqrt($a), sqrt(1-$a))
BenchmarkTools.Trial:
  memory estimate:  0 bytes
  allocs estimate:  0
  --------------
  minimum time:     24.080 ns (0.00% GC)
  median time:      24.870 ns (0.00% GC)
  mean time:        25.987 ns (0.00% GC)
  maximum time:     417.257 ns (0.00% GC)
  --------------
  samples:          10000
  evals/sample:     1000
Main> asin(sqrt(a)) ≈ atan2(sqrt(a), sqrt(1-a))
true
```

* Take minimum after square root in haversine

Author: IlyaOrson

src/haversine.jl

@@ -29,11 +29,8 @@ function evaluate(dist::Haversine, x::VecOrLengthTwoTuple, y::VecOrLengthTwoTupl
     # haversine formula
     a = sin(Δφ/2)^2 + cos(φ₁)*cos(φ₂)*sin(Δλ/2)^2
 
-    # take care of floating point errors
-    a = min(a, one(a))
-
     # distance on the sphere
-    2*dist.radius*atan2(√a, √(1-a))
+    2 * dist.radius * asin( min(√a, one(a)) ) # take care of floating point errors
 end
 
 haversine(x::VecOrLengthTwoTuple, y::VecOrLengthTwoTuple, radius::Real) = evaluate(Haversine(radius), x, y)