fix(docs): sort numeric summary columns containing "varies" sentinel

[bkowshik]

What

Fix numeric sorting on the dataset summary tables (https://moabb.neurotechx.com/docs/dataset_summary.html). Clicking Total_trials (and #Trials / class, #Runs) sorted lexicographically instead of numerically:

Screenshot before the fix, sorted by Total trials (last) column.

before: 11000, 1114, 11496, 1200, 13440, ...
after:  480, 1114, 1200, 1400, ... 250000

Why

The .sortable paradigm tables are initialized with a single shared DataTable() call and no per-column config, so DataTables auto-detects each column's type. Auto-detection is all-or-nothing: a single non-numeric cell downgrades the whole column to string sorting. Several numeric columns carry the varies sentinel for datasets whose trial count isn't fixed, which poisoned detection.

Prior art: #705 fixed this class of bug at the data layer (stripping unit suffixes, e.g. 512Hz -> 512). That approach is fragile — it only holds while every cell stays clean — and #868 reintroduced the bug with varies rows. Unlike a unit suffix, varies carries real meaning and can't just be stripped, so the fix belongs in the sort engine.

Fix

Register a custom DataTables column type num-varies (docs/source/_static/js/sortable-numeric.js) that treats a column as numeric when every cell is a number or a known sentinel (varies, blank, n/a, ...), ordering sentinel rows as -Infinity so they sink to the bottom. Displayed text is unchanged — varies still reads as varies. Wired in via conf.py. No CSV data changed.

Screenshot after the fix, sorted by Total trials (last) column.

Verification

Sort-key logic verified against the real summary_imagery.csv values; docs build green. Final check: click Total_trials in the rendered page.