Preserve fitting post-replacement output during modern truncation

[emme1t]

With the modern algorithm, enabling a length limit can change a slug which already fits that limit:

from slugify import slugify

options = dict(
    algorithm="modern",
    word_boundary=True,
    replacements=[("one", "one---")],
    replacement_stage="post",
)

print(slugify("one two", **options))
# one----two
print(slugify("one two", max_length=100, **options))
# one-two

The README specifies that post replacements remain unfiltered. The word-boundary truncation loop drops empty internal tokens, which collapses intentionally repeated delimiters even when the whole output fits. At an exact length limit, the hard-cut branch can also drop an intentionally added trailing delimiter.

Check the final emitted length before truncating. When the entire output fits, return it with the configured separator mapping intact. The length calculation accounts for empty and multi-character separators without constructing an expanded output just to measure it. This change is confined to the modern helper; the default legacy algorithm and public smart_truncate retain their implementations.

Validation on Windows with Python 3.13.13:

• The original implementation fails 28 new regression subcases covering repeated/boundary delimiters, exact/spare budgets, empty/multi-character separators, word boundaries, and save_order.
• Full suite: 105 passed, 104 subtests passed. This includes the unchanged legacy suite and its 2,688-case differential comparison against the frozen reference.
• Strict Mypy, repository-configured pycodestyle and flake8 commands, and git diff --check all pass.

Dojo follow-up: OpenAI Codex assisted with the investigation, patch, regression tests, and this PR description. The reproductions and checks above were executed locally.