Switch docs theme from alabaster to furo

Reorganize the README's how-it-works prose around the two-layer model:
a vocabulary layer that recognizes pieces by what they are (the
customizable sets), and a positional layer that assigns the rest by
where it sits. States explicitly that parsing is deterministic (no
statistical models) and that the positional layer assumes Western name
order. All claims from the previous prose are preserved.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

These marks (U+200F and friends) ride along with copy-pasted RTL names and
render as nothing, but they glue onto first/last so the parsed parts quietly
stop comparing equal to the clean string. Stripped first, before the rest of
preprocessing, so a mark sitting inside a token can't defeat the Ph.D. and
nickname matching.

Strip invisible bidi control characters during preprocessing