Fix precision loss in Newton-Raphson inverse

[tompng]

The residual precision SIZET2NUM(SIZET2NUM(n / 2)) was a double conversion. It accidentally requested about n digits, and the intended n / 2 is not enough: the digit schedule n = (prec >> i) + 2 leaves only 1-2 margin digits, so rounding errors compound over iterations and the inverse loses many digits. The correction loop in divmod_by_inv_mul then runs once per lost unit of the quotient and a single divmod can take minutes. With ROUND_DOWN or ROUND_FLOOR the same divergence happened even with the accidental n-digit residual, because the iteration converges from below and truncation adds error in the same direction at every step.

• Use 4 margin digits for the schedule and n / 2 + 4 digits for the residual
• Calculate with ROUND_HALF_UP in VpDivdNewton and restore the rounding mode together with the precision limit
• Initial approximation from the leading 18 digits with 64-bit division (about 8 correct digits instead of 2), and skip iterations below that. The previous one could produce frac[0] == BASE for x like 1.001
• Pass VpDivdNewtonInner arguments as a struct

Additional information:
Initial approximation was 2-digits precision because BASE_FIG was 4(uint64 unavailable) or 9(uint64 available) before. 2-digits is the half of BASE_FIG=4.
BASE_FIG is now always 9, so we can increase the initial approximation and the margin.
(Larger margin requires larger initial approximation precision)