FIX: escape literal hyphens in TokenBijectionConverter to avoid delimiter collision

[Kushal Talati (kushaltalati)]

Description

Closes #3128.

TokenBijectionConverter joins the code words of one plaintext word with - and passes every unmapped character through unchanged, so a literal hyphen in the prompt is emitted as the same character as the delimiter and lands exactly where an inserted delimiter would. e-mail and email encode to the same text, and decode() reads both back as email; self-harm comes back as selfharm. The class docstring promises the encoding is unambiguously decodable, and the teaching instructions give the target no way to write a literal hyphen.

This is the third instance of the collision class already fixed in this file for DigitBijectionConverter: #2696 (literal apostrophes vs. _CASE_MARKER) and #2741 (literal digits vs. digit tokens). The fix follows the same pattern:

• encode() writes a literal hyphen as a doubled delimiter (--), which can never be produced by joining code words.
• _decode_pattern gains a leading -- alternative so a doubled delimiter is consumed as one literal hyphen before any code-word run is considered, and decode() restores it.
• get_teaching_instructions() tells the target to write a literal hyphen as two hyphens.
• A comment next to _TOKEN_DELIMITER documents the collision, in the style of the existing _CASE_MARKER / _LITERAL_MARKER comments.

Hyphens that are not directly between two letters (a - b, -lead, trail-) already survived because the pattern requires a code word on both sides; they still round-trip, now via the escape.

Tests and Documentation

Added to tests/unit/converter/test_bijection_converter.py:

• test_token_converter_round_trips_literal_hyphens, parametrized over e-mail, well-known, self-harm, Self-Harm, a - b, -lead, trail-, double--dash, ---.
• test_token_converter_literal_hyphen_is_distinct_from_delimiter: e-mail and email encode differently, and the hyphenated form is the first code word, --, then the rest of the joined form.
• test_token_converter_teaching_instructions_explain_hyphen_escape.

With only the source change reverted, 6 of the new cases fail (the four hyphenated words plus the two new assertions); with it, the file is 151 passed and tests/unit/converter is 2114 passed. ruff check and ruff format --check are clean. The converter notebook in doc/code/converters/1_text_to_text_converters.py uses a prompt without a hyphen, so its output is unchanged and no JupyText run was needed.