Support for PDMats

I would like KeyedArrays to play nice as parameters of Distributions.

This already works in most cases, but there are problems with PDMats.

There is a full interface to define for new subtypes of AbstractPDMat, but I don't know if all of that is warranted.

Constructor example

julia> using AxisKeys, Distributions, PDMats

julia> pdmat = PDMat([1.0 0.0; 0.0 1.0])
2×2 PDMat{Float64,Array{Float64,2}}:
 1.0  0.0
 0.0  1.0

julia> PDMat(KeyedArray([1.0 0.0; 0.0 1.0], ([:x1, :x2], [:y1, :y2])))
ERROR: MethodError: no method matching PDMat(::KeyedArray{Float64,2,Array{Float64,2},Tuple{Array{Symbol,1},Array{Symbol,1}}})
Closest candidates are:
  PDMat(::AbstractArray{T,2} where T, ::LinearAlgebra.Cholesky{T,S}) where {T, S} at /Users/bencottier/.julia/packages/PDMats/Rw2Hf/src/pdmat.jl:12
  PDMat(::LinearAlgebra.Symmetric) at /Users/bencottier/.julia/packages/PDMats/Rw2Hf/src/pdmat.jl:20
  PDMat(::Array{T,2} where T) at /Users/bencottier/.julia/packages/PDMats/Rw2Hf/src/pdmat.jl:19
  ...
Stacktrace:
 [1] top-level scope at REPL[10]:1

This prevents passing a KeyedArray as a covariance matrix in a Distribution.

julia> d = MvNormal(
           KeyedArray([1.0, 2.0], [:a, :b]),
           KeyedArray([1.0 0.0; 0.0 1.0], ([:a, :b], [:a, :b]))
       )
ERROR: MethodError: no method matching PDMats.PDMat(::KeyedArray{Float64,2,Array{Float64,2},Tuple{Array{Symbol,1},Array{Symbol,1}}})
Closest candidates are:
  PDMats.PDMat(::AbstractArray{T,2} where T, ::LinearAlgebra.Cholesky{T,S}) where {T, S} at /Users/bencottier/.julia/packages/PDMats/Rw2Hf/src/pdmat.jl:12
  PDMats.PDMat(::LinearAlgebra.Symmetric) at /Users/bencottier/.julia/packages/PDMats/Rw2Hf/src/pdmat.jl:20
  PDMats.PDMat(::Array{T,2} where T) at /Users/bencottier/.julia/packages/PDMats/Rw2Hf/src/pdmat.jl:19
  ...
Stacktrace:
 [1] MvNormal(::KeyedArray{Float64,1,Array{Float64,1},Base.RefValue{Array{Symbol,1}}}, ::KeyedArray{Float64,2,Array{Float64,2},Tuple{Array{Symbol,1},Array{Symbol,1}}}) at /Users/bencottier/.julia/packages/Distributions/cNe2C/src/multivariate/mvnormal.jl:211
 [2] top-level scope at REPL[92]:1

\ operator example

Another case is the \ operator, which is used by e.g. Distributions._logpdf:

julia> pdmat = PDMat([1.0 0.0; 0.0 1.0]);

julia> pdmat \ [1.0, 2.0]
2-element Array{Float64,1}:
 1.0
 2.0

julia> pdmat \ KeyedArray([1.0, 2.0], [:x, :y])
ERROR: MethodError: \(::PDMat{Float64,Array{Float64,2}}, ::KeyedArray{Float64,1,Array{Float64,1},Base.RefValue{Array{Symbol,1}}}) is ambiguous. Candidates:
  \(a::PDMat, x::Union{AbstractArray{T,1}, AbstractArray{T,2}} where T) in PDMats at /Users/bencottier/.julia/packages/PDMats/Rw2Hf/src/pdmat.jl:50
  \(x::AbstractArray{T,2} where T, y::KeyedArray{T,1,AT,RT} where RT where AT where T<:Number) in AxisKeys at /Users/bencottier/.julia/packages/AxisKeys/1jgJz/src/functions.jl:303
Possible fix, define
  \(::PDMat, ::KeyedArray{T,1,AT,RT} where RT where AT where T<:Number)
Stacktrace:
 [1] top-level scope at REPL[12]:1

julia> d = MvNormal(KeyedArray([1.0, 2.0], [:a, :b]), [1.0 0.0; 0.0 1.0])
MvNormal{Float64,PDMats.PDMat{Float64,Array{Float64,2}},KeyedArray{Float64,1,Array{Float64,1},Base.RefValue{Array{Symbol,1}}}}(
dim: 2
μ: [1.0, 2.0]
Σ: [1.0 0.0; 0.0 1.0]
)

julia> logpdf(d, [1.0, 2.0])
ERROR: MethodError: \(::LinearAlgebra.LowerTriangular{Float64,Array{Float64,2}}, ::KeyedArray{Float64,1,Array{Float64,1},Base.RefValue{Array{Symbol,1}}}) is ambiguous. Candidates:
  \(A::Union{LinearAlgebra.LowerTriangular, LinearAlgebra.UpperTriangular}, B::AbstractArray{T,1} where T) in LinearAlgebra at /Applications/Julia-1.5.app/Contents/Resources/julia/share/julia/stdlib/v1.5/LinearAlgebra/src/triangular.jl:2050
  \(x::AbstractArray{T,2} where T, y::KeyedArray{T,1,AT,RT} where RT where AT where T<:Number) in AxisKeys at /Users/bencottier/.julia/packages/AxisKeys/1jgJz/src/functions.jl:303
Possible fix, define
  \(::Union{LinearAlgebra.LowerTriangular, LinearAlgebra.UpperTriangular}, ::KeyedArray{T,1,AT,RT} where RT where AT where T<:Number)
Stacktrace:
 [1] invquad(::PDMat{Float64,Array{Float64,2}}, ::KeyedArray{Float64,1,Array{Float64,1},Base.RefValue{Array{Symbol,1}}}) at /Users/bencottier/.julia/packages/PDMats/Rw2Hf/src/pdmat.jl:79
 [2] sqmahal(::MvNormal{Float64,PDMat{Float64,Array{Float64,2}},KeyedArray{Float64,1,Array{Float64,1},Base.RefValue{Array{Symbol,1}}}}, ::Array{Float64,1}) at /Users/bencottier/.julia/packages/Distributions/cNe2C/src/multivariate/mvnormal.jl:266
 [3] _logpdf(::MvNormal{Float64,PDMat{Float64,Array{Float64,2}},KeyedArray{Float64,1,Array{Float64,1},Base.RefValue{Array{Symbol,1}}}}, ::Array{Float64,1}) at /Users/bencottier/.julia/packages/Distributions/cNe2C/src/multivariate/mvnormal.jl:127
 [4] logpdf(::MvNormal{Float64,PDMat{Float64,Array{Float64,2}},KeyedArray{Float64,1,Array{Float64,1},Base.RefValue{Array{Symbol,1}}}}, ::Array{Float64,1}) at /Users/bencottier/.julia/packages/Distributions/cNe2C/src/multivariates.jl:201
 [5] top-level scope at REPL[18]:1

Start with AxisKeys/src/functions.jl at the keyed-array \ method and reproduce the constructor, \, and MvNormal/logpdf examples from the issue. Read the referenced PDMats pdmat.jl and Distributions mvnormal.jl call paths to understand the method ambiguities. Done means the shown KeyedArray covariance and vector cases construct successfully and the MvNormal logpdf example runs without ambiguity.