[dev-rinchin]

Compiled CPU inference can return different predictions for equivalent sparse matrices, or after enabling quantize=1. For example, a CSR coordinate stored as 0.4 + 0.4 follows the wrong branch at a 0.5 split, and quantization changes one prediction in a 128-tree regression fixture by up to 0.08035.

This fixes both cases while retaining float32 inference and threshold quantization. Follow-up to #37.

• Sum duplicate sparse coordinates in the input dtype before casting to float32 or dropping NaNs. Copy noncanonical CSR buffers before summation, preserve explicit zeros, and leave caller-owned matrices unchanged.
• Preserve split boundaries affected by TL2cgen 1.0's eight-digit float-array formatting. On a separate compilation copy, rewrite an affected x <= t comparison as the equivalent float32 comparison x < nextafter(t, +inf). Apply this at compile time so existing Treelite 4 checkpoints also benefit; keep the original model and compiler parameter dictionary unchanged.
• Add 40 regression cases covering CSR/CSC/COO, float32/float64 values, int32/int64 indices, cancellation, NaN duplicates, threshold equality and neighboring values, both NaN directions, scalar/vector outputs, checkpoint loading, and serial/parallel compilation. Update the README with the supported behavior.

Validation:

• Full Linux/NVIDIA A100 suite on the resulting tree: 419 passed, 1 Windows-only test skipped.
• Built sdist and wheel from the final branch; installed the wheel and ran the full suite outside the source checkout without CuPy: 116 passed, 206 skipped. GPU tests and the actual Windows loader test are skipped in this environment.
• CPU integration and loader suites on Python 3.8 / NumPy 1.24.4 / Treelite 4.1.2, Python 3.12 / Treelite 4.1.2, and Python 3.14 / Treelite 4.8.1: 110 passed, 1 skipped in each environment, with TL2cgen 1.0.0.
• Original 128-tree fixture: quantized maximum absolute error falls from 0.08035 to 3.8e-6, matching ordinary compilation. An additional probe of 64 affected float32 thresholds across signs and exponents passes for equality, neighboring values, NaNs, and infinities.
• Ruff syntax/undefined-name checks, uv pip check, and git diff --check pass. Local Windows execution is not available; the existing CI matrix covers GCC and MSVC.
• GitHub Actions: all 8 Linux/Windows jobs passed, including GCC/MSVC and the minimum Treelite version. Successful CI run.