AdvancedPS v0.7 (and thus Libtask v0.9) support

[mhauru]

The complement PR of TuringLang/AdvancedPS.jl#114, which adds support for the newly rewritten Libtask.

Work in progress, currently blocked by TuringLang/Libtask.jl#186

[mhauru]

The tests that I had the patience to run locally now pass. Waiting for the AdvancedPS release to be able to run the full test suite on CI.

Some indicators of speed:

julia> module MWE

       using Turing

       @model function gdemo(x, y)
           s ~ InverseGamma(2, 3)
           m ~ Normal(0, sqrt(s))
           x ~ Normal(m, sqrt(s))
           y ~ Normal(m, sqrt(s))
           return s, m
       end

       @time chn = sample(gdemo(2.5, 1.0), PG(10), 10_000)
       describe(chn)

       end

On main:

104.715858 seconds (58.48 M allocations: 13.259 GiB, 1.10% gc time, 1.20% compilation time)

On this branch:

 16.612050 seconds (116.52 M allocations: 16.296 GiB, 8.27% gc time, 5.91% compilation time)

julia> module MWE

       using Turing

       @model function f(dim=20, ::Type{T}=Float64) where T
           s = Vector{Bool}(undef, dim)
           x = Vector{T}(undef, dim)
           for i in 1:dim
               s[i] ~ Bernoulli()
               if s[i]
                   x[i] ~ Normal()
                else
                   x[i] ~ Beta()
                end
                0.0 ~ Normal(x[i])
           end
           return nothing
       end

       alg = Gibbs(
           @varname(s)=>PG(10),
           @varname(x)=>HMC(0.1, 5),
       )
       @time chn = sample(f(), alg, 1_000)

       end

On main:

 49.682945 seconds (65.43 M allocations: 9.463 GiB, 1.79% gc time, 8.23% compilation time)

On this branch:

  9.180071 seconds (61.58 M allocations: 4.028 GiB, 4.04% gc time, 55.49% compilation time)

Obviously the speed gains are all due to @willtebbutt's fantastic work on Libtask, everything else is just wrapping that work.