Fix size_hint underflow on empty-string sentence iterators

[lenamonj]

size_hint() on any of the three public sentence iterators panics on the empty string in debug builds:

"".split_sentence_bounds().size_hint();       // attempt to subtract with overflow
"".split_sentence_bound_indices().size_hint(); // same
"".unicode_sentences().size_hint();            // same

In release builds the same calls return (18446744073709551615, Some(0)) - a lower bound above the upper bound, which the Iterator::size_hint contract forbids. A consumer that passes the lower bound to Vec::with_capacity gets a capacity-overflow panic; that appears to be exactly what the reporter of #146 ran into in a comment on that thread ("something in that logic managed to summon a value that made Vec::with_capacity() panic with a Capacity overflow by just passing the lower bound unaltered") - the issue was closed on the min misreading, but this underflow was live underneath it.

Cause: USentenceBounds::size_hint subtracts 1 from the inner iterator's lower bound, and that bound is 0 for an empty string. The subtraction is on usize, and the cmp::max(0, ...) wrappers around both bounds are no-ops on an unsigned type that read as saturation while the underflow sat inside them.

Fix: both bounds use saturating_sub(1); the wrappers and the now-unused top-level use core::cmp are removed. For every non-empty input the arithmetic is unchanged, since cmp::max(0, x - 1) and x.saturating_sub(1) agree wherever x >= 1; only the empty-string case moves, from a panic (debug) or a wrapped bound (release) to (0, Some(0)).

Test: the new test_size_hint_is_a_valid_bound drives size_hint on all ten public iterators over a corpus led by the empty string plus the crate's own test tables, asserting lower <= yielded <= upper. On master it fails with the debug panic above; with the fix, cargo test, cargo fmt --check, cargo clippy --all-targets --all and cargo +1.85.0 test (the MSRV job's command) all pass.

One related observation, kept out of this PR to keep it minimal: the inner SentenceBreaks::size_hint derives its bounds from the whole string's length rather than the unconsumed remainder, so the outer hint never shrinks as items are yielded (a lower bound of 1 remains after the last item). Happy to follow up separately if that is of interest.

---

Provenance, for transparency: this defect was found by an automated audit loop I run against open-source projects; the patch and its verification were reviewed by me before filing, and I am happy to answer any questions during review.