Rewrite optimal FLCI with first-difference reparameterization

The previous FLCI bias computation was wrong: it either pinned
delta_pre=0 (making bias zero) or left levels unconstrained (making
bias infinite). The paper (Section 4.1.1) parameterizes the estimator
in terms of weights on pre-treatment first differences (slopes):

  theta_hat = l'beta_post - sum_s w_s (beta_s - beta_{s-1})

with constraint sum(w)=1 (linear trend invariance). The bias LP
operates in first-difference space where Delta^SD is a bounded
polyhedron |fd_{i+1} - fd_i| <= M, making the LP well-posed.

New helper functions:
- _build_fd_transform: maps first-differences to levels (C matrix)
- _build_fd_smoothness_constraints: Delta^SD in fd-space
- _w_to_v: maps slope weights w to estimator direction v

_cv_alpha now supports optional df parameter using scipy.stats.nct
(folded non-central t) for survey inference on the smoothness path.

M=0 naturally falls out: bias=0 when sum(w)=1, optimizer just
minimizes variance, giving exact linear-extrapolation weights
including pre-period contribution to SE.

87/87 tests pass. New regressions: survey df=2 widens smoothness CI,
M=0 SE includes pre-period variance, bias nonzero for M>0.

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

Author: igerber