massively reduce CPU allocation

[Moelf]

setup:

julia> x_train = rand(10^4, 60);

julia> y_train = rand(size(x_train, 1));

julia> config = EvoTreeRegressor(
               loss=:logistic,
               nrounds=100,
               lambda=0.5, gamma=0.0, eta=0.05,
               max_depth=6, min_weight=1.0,
               rowsample=1.0, colsample=0.5, nbins=64);
[ Info: Following 5 arguments were not provided and will be set to default: alpha, T, rng, device, monotone_constraints.

Before

julia> @time fit_evotree(config; x_train, y_train);
  0.330572 seconds (440.90 k allocations: 218.696 MiB, 9.79% gc time)

After this PR

julia> @time fit_evotree(config; x_train, y_train);
  0.242588 seconds (343.28 k allocations: 54.385 MiB, 6.30% gc time)

[codecov-commenter]

Codecov Report

Base: 50.36% // Head: 50.56% // Increases project coverage by +0.19% 🎉

Coverage data is based on head (3a18d0c) compared to base (875e93f).
Patch coverage: 100.00% of modified lines in pull request are covered.

Additional details and impacted files

@@            Coverage Diff             @@
##             main     #186      +/-   ##
==========================================
+ Coverage   50.36%   50.56%   +0.19%     
==========================================
  Files          17       17              
  Lines        1487     1493       +6     
==========================================
+ Hits          749      755       +6     
  Misses        738      738              

Impacted Files	Coverage Δ
src/fit.jl	94.37% <100.00%> (+0.21%)	⬆️

Help us with your feedback. Take ten seconds to tell us how you rate us. Have a feature suggestion? Share it here.

☔ View full report at Codecov.
📢 Do you have feedback about the report comment? Let us know in this issue.

[Moelf]

@jeremiedb btw, I think for all the Vector{Vector{}} stuff, you should use https://github.com/JuliaArrays/ArraysOfArrays.jl

its supports syntax of Vector{Vector{}} while also providing flattenview which can help you quickly updates all the underlying values in node (without pointer chasing) and also GPU friendly.