Fixes to max_daf returns, under various "empty" scenarios - #396

The scalar max_daf kept a site with as little as one call (n_valid > 0),
while both windowed engines already required a site's own pair (n_valid
>= 2) -- the package convention from that pi/theta_w/
heterozygosity_expected also follow, since a frequency estimate from one
call is not useful. A single stray derived-allele call could set the
whole-region scalar value to that call's own frequency (up to 1.0) while
the windowed value, on the same data, correctly ignored it.

Bumped the scalar's include-mode gate to n_valid >= 2 to match. Docstring
now says so explicitly.

Tests: the issue's own reproducer, pinned to the windowed value (1/3, not
1.0), plus a direct comparison against both the scatter and fused
engines on the same data. Checked the existing missing-data parity
fixture in test_implementation_parity.py always keeps at least two valid
calls per site, so it doesn't exercise this gate either way and needed no
change. Also updates the now-stale pipeline_map.toml note (it previously
documented the scalar/windowed mismatch as a known fact) and the
changelog.

The scalar max_daf(missing_data='exclude') now gives NaN when no site is
complete (previous commit). The windowed engines had the same gap one
level up: _windowed_thetas_scatter, windowed_statistics_fused, and
windowed_statistics_fused_chunked all filter the whole matrix to complete
sites once, upfront, before windowing even starts. By the time a window's
variant count is checked, a window that originally held variants but lost
them all to exclusion is indistinguishable from a window that never had a
variant to begin with -- both just have zero post-filter variants, and
both engines left max_daf at its zero-initialized default either way. The
first case is undefined (nothing to measure); the second is the
established "0.0 for an empty window" convention every additive statistic
already uses, which does NOT apply to a max.

Fixed by capturing each window's variant count from the ORIGINAL
(pre-filter) positions before exclusion runs, alongside the existing
post-filter count: a window with pre-filter variants but zero post-filter
ones gets NaN, mirroring the existing zero_span post-hoc override pattern
in the scatter engines (compute a per-window flag once, apply it via
np.where at the point the result is built) rather than threading a
special case through the max-accumulation logic itself. Needed in all
three engines, including the chunked one, since chunking only changes how
the kernel is invoked, not how the window-to-variant-index ranges are set
up before it.

Tests: a two-window matrix (one window has two sites each with a missing
call -- wiped out by exclude; the other has no variant at all) run
through all three engines plus the scalar function on the equivalent
subset, confirmed to fail without the fix. Also updates the
pipeline_map.toml undefined/notes text and the changelog.