Fix line number recognition for sparse and accumulated numbering

[lfoppiano]

Summary

TextPage::markLineNumber() misses valid line-number columns in two layouts the size-scaled
coverage threshold from #226 does not cover — and a naive fix wrongly strips footnote/reference
numbers (which also accumulate to large values at a margin). This recovers the missed line numbers
while leaving footnote/reference numbers intact, using a document-level guard.

Detection (low-coverage escape hatch)

Accept a margin cluster as line numbers when its values form a regular arithmetic progression
(dominant constant increment) at the extreme page margin, via either:

• big numbers — the smallest value exceeds the page's line count: accumulated numbering that a
  figure or column break may split (e.g. 428..444 on a late page), or
• even vertical pitch — fresh sparse numbering such as every 5th line (229336v1: coverage
  ~0.17 vs the required 0.5, so 242 margin numbers used to survive).

Footnote guard (document level)

Footnote / bibliographic reference numbers also accumulate to large values at a margin, but their
block sits at the bottom of the page and never covers it, whereas a line-number column reaches
the top of the page's text on some page and runs down every page. So a low-coverage column is
accepted only if it reaches the top of the text here, or was already confirmed as a line-number
column on an earlier page (2-column pages and figures push the numbers down, so a confirmed
column may legitimately start mid-page). Confirmed column x-positions are carried across pages via
XmlAltoOutputDev. Per-page geometry alone cannot separate the two — a figure above the text
pushes line numbers into the same vertical band as footnotes; the document-level view does.

Validation

Measured on the GROBID end-to-end corpora and the grobid-footnote-flavour corpus, diffing
pdfalto output unpatched vs patched.

Recall — line numbers stripped (ALTO token level):

corpus	docs	docs stripped	tokens	over-strip	negative deltas
bioRxiv-10k-test-2000	2000	468 (23%)	17,570	0	0
bioRxiv-10k-train-6000	6000	1,655 (28%)	59,165	0	0

Only digit tokens are removed (verified token-by-token on the heaviest strips). Vs the un-guarded
version, 94% of recall retained on test-2000 and 97% on train-6000; the residual are
2-column/figure docs whose line-number column never reaches the page top on any page (inherently
ambiguous with footnotes).

Precision — nothing that isn't a line number is removed:

Stress-tested against the full mpilhlt/grobid-footnote-flavour
corpus — 100 two-column, footnote-heavy legal/humanities PDFs (all 12 batches), the exact layout
where footnote/reference numbers accumulate to large values at a margin and look like line numbers:

corpus	docs	docs changed (before guard → after)
grobid-footnote-flavour (all batches)	100	2 → 0 — no footnote/page numbers stripped

• eLife/PLOS/PMC (no line numbers): 0 changed, unchanged.
• Figure-panel labels (005124v1, inset 1..6) preserved.

Notes

Conservative constants: increment 1–50, pitch CV < 0.30, margin within 6% of page width, "reaches
top" = topmost number within the top 40% of the page's text, confirmed-column x tolerance 15 pt.