Add optimize() method for automatic layout and chunk_size tuning

[dantegd]

PR ports the optimize() method from cuML FIL to cuforest, enabling automatic tuning of layout and chunk_size parameters for optimal inference performance.

Changes

• Mutable layout support: Models can now change layout after construction, triggering automatic model reload
• optimize() method: Searches over layout (depth_first, breadth_first, layered) and chunk_size (powers of 2) combinations to find optimal configuration for a given batch size
• Settable default_chunk_size: Allows setting the default chunk size used when none is specified in predict calls

Files Modified

File	Changes
detail/forest_inference.pyx	Store treelite bytes for reload, add layout setter with _reload_model()
_forest_inference.py	Add OptimizeMixin class with optimize() method
_base.py	Add abstract signatures for optimize(), layout setter, default_chunk_size
tests/test_optimize.py	New test file with 7 tests for optimize functionality

Usage

import cuforest

fm = cuforest.load_model("model.ubj", device="gpu")

# Auto-tune for a specific batch size
fm.optimize(data=X, timeout=0.5)

# Or manually set layout/chunk_size
fm.layout = "breadth_first"
fm.default_chunk_size = 16

[copy-pr-bot]

This pull request requires additional validation before any workflows can run on NVIDIA's runners.

Pull request vetters can view their responsibilities here.

Contributors can view more details about this message here.

[hcho3]

/ok to test 9ec0856

[hcho3]

/merge

[csadorf]

Post-merge review; I have one high-level concern.

[csadorf]

I think we should revise the model where we optimize an instance "in-place" that necessitates some of the complexity here for "reloading" a model. We should instead consider whether we simply produce a new "optimized" instance.

[dantegd]

Good point. The in-place mutation with hidden model reload on layout change is surprising behavior. I ca revise to have optimize() return a new instance with optimal settings instead:

fm_opt = fm.optimize(data=X)