FIX: rejoin words with the configured word_split_separator

[Umerkhan Golandaz (UmerkhanGolandaz)]

Description

Fixes #2592.

WordLevelConverter.convert_async splits the prompt on word_split_separator (word_level_converter.py:116), but the default join_words always rejoined with a space (:91). The split and the join disagreed, so a custom separator was silently replaced in the output:

EmojiConverter(seed=1, word_split_separator=",").convert_async(prompt="alpha,beta,gamma")
# before: 🅐🅛🄿🅗🄰 🄱🅴🅣🅰️ 🄶🅐🅜🄼🅰️   <- commas gone
# after:  🅐🅛🄿🅗🄰,🄱🅴🅣🅰️,🄶🅐🅜🄼🅰️   <- preserved

EmojiConverter and BinAsciiConverter both expose word_split_separator in their own constructors, so this is reachable through the public API.

Changes

• join_words rejoins with the separator the words were split on.
• A None separator splits on arbitrary whitespace, which has no single representation to restore, so it continues to join with a space.
• Tests covering custom separators (,, |, -, ::), the unchanged default, the None case, and a custom separator combined with partial word selection.

Why this is safe

The change is a no-op for every caller that does not set a custom separator:

• word_split_separator=" " — the default — already joined with " "
• word_split_separator=None — still joins with " "

Only the custom-separator path changes, and it is currently broken. No existing test asserted the old behavior: tests/unit/converter/test_emoji_converter.py:53 sets word_split_separator="|" but asserts only the identifier params, not the output.

Subclasses that override join_words (BinaryConverter, NatoConverter, FirstLetterConverter, UnicodeReplacementConverter, BinAsciiConverter) keep their own joining behavior. BinAsciiConverter delegates to super().join_words outside its all-words encodings, so it picks up the fix on that path.

Tests

• pytest tests/unit/converter/ — 1445 passed, 34 skipped, including the 5 new tests
• ruff format --check and ruff check clean on both changed files