Token grouping stage: collapses opaque regions (reader conditionals,
namespaced maps, syntax-quote with brackets) from flat token sequences
into single composite tokens.

This replaces the tokenizer's read-balanced-raw by operating on
already-tokenized input where bracket matching is trivial — strings,
chars, and comments are already individual tokens.

Scanner-level source-position utilities.
Defines the character-level line/col model used by the tokenizer and grouper.
Only \n advances the line counter — \r is a regular character that occupies
a column. This matches sadvance! in the tokenizer.

Note: this is the *scanner* line model, not a universal line definition.
The error display module (beme.alpha.errors/source-context) uses
str/split-lines which has different line-ending semantics (splits on
both \n and \r\n). The two models agree for LF sources but diverge
for CRLF. See format-error for how the bridge is handled.

beme tokenizer: character scanning and token production.
Transforms beme source text into a flat vector of typed tokens.