Trigonometric argument reduction for float16

[pearu]

This PR adds:

• split_veltkamp2(x) to be used for large inputs,
• Pade and Taylor approximations for sine and cosine to be used for small inputs,
• series(x, y, ...) expression kind to represent series x * 2 ** sexp + y * 2 ** (2*sexp) + ...,
• addition, subtraction, multiplication and division of series using native/2sum and/or native/dekker multiplication,
• fast_polynomial2 and fast_polynomial_dekker for evaluating polynomials,
• fma for computing multiply-add. Fused multiply-add is emulated with upcasting of arguments and downcasting of a result,
• sine_taylor_dekker and cosine_taylor_dekker to be used for small inputs. These functions provide best accuracy,
• sine on the whole real line.