MPSBoost histogram benchmark

Historical baseline note: these measurements were recorded before the later v2/v3 optimization and estimator cleanup work. They are retained for traceability and should not be read as current HEAD or final 0.3.0 performance claims.

Wall time includes host preparation, buffer transfer, command submission, and synchronization.

Scenario	Rows	Features	CPU median (s)	MPS median (s)	Speedup	Pool reuse
small	4096	16	0.000736	0.001613	0.456x	6
medium	32768	32	0.005142	0.004310	1.193x	13
large	131072	64	0.035795	0.012662	2.827x	20
wide	32768	256	0.042130	0.015056	2.798x	27

MPSBoost end-to-end regressor benchmark

Historical baseline note: these measurements were recorded before the later v2/v3 optimization and estimator cleanup work. They are retained for traceability and should not be read as current HEAD or final 0.3.0 performance claims.

Wall time includes Python input adaptation, quantization, training, model assembly, and synchronization.

Scenario	Rows	Features	CPU median (s)	MPS median (s)	Speedup	Max prediction diff
gbdt-medium	16384	32	0.067630	0.089597	0.755x	3.36394e-06
gbdt-wide	16384	128	0.283176	0.249502	1.135x	4.35114e-06
gbdt-large-wide	32768	256	1.031002	0.633006	1.629x	5.36442e-06