feat(thresholds): add geometric endemic channel strategy

[knutdrand]

What

Adds a geometric endemic channel strategy alongside the existing seasonal and percentile ones:

threshold = expm1(mean(log1p(cases)) + k * std(log1p(cases)))

over historical same-month (or same-week) values — the geometric mean times the geometric standard deviation raised to the power k. std_multiplier takes a scalar or a list, so one request can return several lines.

Why

Case counts are strongly right-skewed, and seasonal's arithmetic mean is pulled up by single epidemic years. Working on a log scale centres the channel on the geometric mean instead, which AM-GM keeps at or below the arithmetic mean for any non-constant history.

Computing on log1p/expm1 rather than a plain geometric mean is deliberate: it keeps zero counts, which a plain geometric mean cannot represent, and an all-zero season yields a finite threshold of 0 rather than dropping the rows.

Caveat, stated in the strategy description

Only the centre is guaranteed lower than the arithmetic channel. The band is multiplicative, so the log-scale standard deviation is a relative spread, and on a short baseline a single extreme year can make the back-transformed band wider than mean + k*std:

baseline (one 20x spike)	seasonal	geometric
n=5	2179	2652 (wider)
n=8	1681	1211
n=20	1045	445

That matters because percentile defaults to a 5-year baseline, which is exactly where the effect reverses. The registry description, module docstring and contributor guide all say this, and the test asserts the guaranteed property (k=0, AM-GM) rather than the data-dependent inequality.

Notes

• No endpoint changes — registration through @threshold plus membership in the ThresholdParams union is enough.
• Contributor guide gains a built-in strategy table covering all three strategies and when to reach for each.
• An earlier revision of this branch also added a median_iqr (Tukey fence) strategy; it was dropped as redundant next to percentile.

Testing

• make lint clean (ruff, mypy, pyright)
• Full suite: 1501 passed, 122 skipped, 4 xfailed
• New tests cover values, parity with the module-level helper, zero counts, the AM-GM centre against an epidemic-year baseline, multi-line output, weekly data, frequency mismatch, and params validation, plus an end-to-end endpoint test