[FEA] Roadmap for paying down tech debt in Treelite

[hcho3]

cuML uses dmlc/treelite to implement FIL and tree serialization. Treelite currently has many outstanding issues:

• Model serialization is slow (Serialization of large trees to protobuf is slow - consider flat representation dmlc/treelite#156)
• LightGBM and sklearn models require float64, but Treelite uses float32 (In our experiment, origin lightgbm predict result is not equal treelite predict result dmlc/treelite#77 my litghtgbm model can't get correct predict value  dmlc/treelite#94 predict result is different, after compile with treelite dmlc/treelite#155)
• Treelite doesn't support majority voting for multi-class classifiers ([Draft] Improved model specification, to clarify how multiclass classification is handled dmlc/treelite#149)
• cuML uses Conda to set up prerequisites and dependencies. Treelite is a lone exception and not available on Conda.
• The Treelite package on PyPI is very outdated.
• CMake build and setup.py in Treelite is quite complex and broken in some places. For example, the Treelite runtime has its own CMakeLists.txt that copies headers from the parent Treelite project at CMake configuration stage. Running make clean will remove the copied headers and break the build.

To resolve these issues, I will want to do a whole-sale refactoring of Treelite. The refactoring will be mostly on the build systems and the serialization.

I'm opening this issue to keep track of progress.

• Release Treelite 0.90 using the latest master (Release 0.90 dmlc/treelite#164)
• Re-design CMake build system using Modern CMake principles (Refactor CMake build; create treelite_runtime Python pkg dmlc/treelite#167)
• Revamp Python tests using Pytest (Refactor Python tests with Pytest dmlc/treelite#176)
• Enable fast binary serialization (Fast binary serialization of Treelite models dmlc/treelite#178, [REVIEW] Faster Treelite serialization #2263)
• Create installation target for Treelite so that other C++ applications can find it using find_package(Treelite) (Add install target in CMake dmlc/treelite#182)
• Submit a Conda recipe to conda-forge (Add treelite recipe conda-forge/staged-recipes#11926)
• Add support for float64

[hcho3]

dmlc/treelite#167 has revamped CMake build system for Treelite.

[hcho3]

Treelite is now officially on Conda: https://anaconda.org/conda-forge/treelite. Install it by running

conda install -c conda-forge treelite=0.92

[hcho3]

Float64 support is added in dmlc/treelite#196. I plan to break up this pull request into smaller ones for easy reviews.

[oreo-lp]

how about float16?