Use Bessel.jl for performance?

[aplavin]

I noticed that the implementations of Bessel functions used here are quite slow, and replacing them with Bessel.jl can give major speedups.
For examples, replacing

Distributions.jl/src/univariate/continuous/rician.jl

Lines 112 to 116 in c2a7387

	function logpdf(d::Rician, x::Real)
	ν, σ = params(d)
	result = log(2 * abs(x) / σ^2) + logpdf(NoncentralChisq(2, (ν / σ)^2), (x / σ)^2)
	return x < 0 || isinf(x) ? oftype(result, -Inf) : result
	end

with

julia> import Bessels as B

julia> @eval Distributions function logpdf(d::Rician, x::Real)
           ν, σ = params(d); σ² = σ^2
           result = log(x/σ² * $B.besseli0(x*ν/σ²)) - (x^2 + ν^2)/2σ²
           return x < 0 || isinf(x) ? oftype(result, -Inf) : result
       end

leads to ~x3-4 performance improvement immediately. Presumably, there are much more cases where this is possible.

Would adding a Bessels.jl dep be considered a reasonable thing by Distributions.jl devs?

[devmotion]

Improving performance of Bessel's function is something to be discussed in SpecialFunctions IMO: JuliaMath/SpecialFunctions.jl#409 Bessels.jl doesn't seem to be without issues either, see e.g. JuliaMath/Bessels.jl#119.