Fix analytic bs_vanna: spurious sqrt(t) factor and sign error

[gaoflow]

bs_vanna returns the analytic Black-Scholes vanna scaled by -sqrt(t). It computes exp(-q t) * sqrt(t) * n(d1) * d2/v, but the correct vanna (∂vega/∂spot = ∂delta/∂vol, the DvegaDspot from #89, added in #94) is -exp(-q t) * n(d1) * d2/v. So there are two compounded errors: a spurious sqrt(t) factor and a flipped sign. The public EquityVanillaOption.vanna() calls bs_vanna directly and inherits both.

I checked the whole analytic greek surface against a closed-form reference and finite-differences of the price: delta, gamma, vega, theta and rho all agree to ~1e-6; only vanna diverged. The base-class EquityOption.vanna (finite-difference of delta) is fine, so the bug is confined to the analytic bs_vanna.

The existing regression only logged vanna into the golden file, which froze the wrong value (e.g. the call vanna was negative for an OTM strike, where it should be positive). This PR adds a real assertion in unit_tests pinning vanna() against both the closed form and the finite-difference of delta w.r.t. vol, and updates the golden vanna column.

Repro at S=100, K=120, t=2, r=0.03, q=0.01, v=0.10:

bs_vanna         -> -3.84489
closed form      ->  2.71875
FD d(delta)/dvol ->  2.71875