Spurious gap before a fused `'s_...` run (genitive/contraction 's split from its host word)

Related to #8 (spurious whitespace around quotes/colons/periods) and to the
underscore-fusion convention used throughout this corpus for multi-word
proper nouns and idioms (e.g. `Fulton_County_Grand_Jury`).

When a genitive or contraction `'s` gets fused via underscore into a larger
compound token together with the word that *follows* it (instead of staying
flush with the word it actually belongs to), the corpus keeps a stray space
or underscore immediately before the `'s`, e.g.:

- `Al 's_Little_Cafe` for Brown's `Al's Little Cafe`
- `Breed 's_Hill` for Brown's `Breed's Hill`
- `Fulton_Tax_Commissioner 's_Office` (internally, one token) for Brown's
`Fulton Tax Commissioner's Office`
- `He 's_got` for Brown's `He's got`

`'s` is never preceded by whitespace in real English (genitive or
contraction), so this is a purely structural bug: whenever a token's text
contains `'s_` preceded by a literal space or underscore, that one
character should be deleted (making `'s` flush with its host word), not
kept as a word-separator. This is unrelated to whether the fused run is its
own separate token or sits in the middle of a larger single token (e.g. the
`Fulton_Tax_Commissioner_'s_Office` case above is one token throughout).

At least one instance (`" Hector 's_Farewell_to_Andromache "` in
`data/belles_lettres/br-g31.yaml`) also blocks an otherwise-correct
quote-gap fix (#8): the spurious gap before `'s` breaks the context match
needed to confirm the opening quote should be flush with `Hector`.

A scan of `data/*.yaml` for `[ _]'s_` (verified against
`src/semcor/brown-nolines.txt`) finds **31 confirmed instances**, all
verified against the reference text.

Start by scanning data/*.yaml for [ _]'s_ and compare the matches with src/semcor/brown-nolines.txt, including the example in data/belles_lettres/br-g31.yaml. Done means all 31 confirmed instances have the stray character before 's removed while the corpus remains aligned with the reference text.

Written by the indexing model from the issue text.