migrate to pybind11

[jmoralez]

Migrates from ctypes to pybind11. The main motivation is to reduce the function call overhead, since for the single array functions it was noticeable, for example the exponentially_weighted_mean function is 2.4x faster with pybind11.

This also allows:

• Using more C++ by not having to expose a C API.
• More safety when passing arguments from Python to C++ (type checks and coercion if necessary).
• More flexibility in function returns, for example we can allocate objects on the C++ side and return them back to python.

[codspeed-hq]

CodSpeed Performance Report

Merging #48 will improve performances by ×2.4

_{Comparing pybind11 (0276b85) with main (f15816a)}

Summary

⚡ 3 improvements
✅ 17 untouched benchmarks

Benchmarks breakdown

	Benchmark	main	pybind11	Change
⚡	test_ewm	140.8 µs	58.8 µs	×2.4
⚡	test_lag_transforms[seasonal_rolling_max]	43.8 ms	38.1 ms	+15.1%
⚡	test_lag_transforms[seasonal_rolling_min]	44.9 ms	39.2 ms	+14.66%