Use ITP as root-finding algorithm in planar layer

[devmotion]

As recommended in EnzymeAD/Enzyme.jl#2035 (comment) for increased robustness.

ITP was added in Roots 1.3.15: JuliaMath/Roots.jl#274

Edit: A quick comparison:

julia> using Bijectors

julia> x = randn(); y = expm1(randn()); z = randn();

julia> Bijectors.find_alpha(x, y, z)
0.20145381857054057 # PR
0.20145381857054054 # master

On the master branch:

julia> using Bijectors, Chairmarks

julia> @be (randn(), expm1(randn()), randn()) Bijectors.find_alpha(_[1], _[2], _[3])
Benchmark: 3014 samples with 15 evaluations
 min    1.350 μs
 median 2.031 μs
 mean   2.052 μs
 max    3.578 μs

julia> @be (randn(), expm1(randn()), randn()) Bijectors.find_alpha(_[1], _[2], _[3])
Benchmark: 3440 samples with 13 evaluations
 min    1.327 μs
 median 2.053 μs
 mean   2.078 μs
 max    9.151 μs

On this PR:

julia> using Bijectors, Chairmarks

julia> @be (randn(), expm1(randn()), randn()) Bijectors.find_alpha(_[1], _[2], _[3])
Benchmark: 2969 samples with 34 evaluations
 min    686.294 ns
 median 890.941 ns
 mean   909.760 ns
 max    1.614 μs

julia> @be (randn(), expm1(randn()), randn()) Bijectors.find_alpha(_[1], _[2], _[3])
Benchmark: 3790 samples with 26 evaluations
 min    663.462 ns
 median 910.250 ns
 mean   929.133 ns
 max    3.385 μs

[yebai]

Thanks @devmotion -- this looks good. The failing tests are due to Enzyme's lack of support for Julia 1.11.

@penelopeysm can you help separate Enzyme tests into a separate CI task and allow it to fail for PRs?