Fix panic in Strip decoder on short all-content tokens

[v-code01]

Problem

Strip::decode_chain's trailing-strip loop panics on short tokens made up of the content char:

let mut stop_cut = chars.len();
for i in 0..self.stop {                 // not bounded by chars.len()
    let index = chars.len() - i - 1;    // underflows once i >= chars.len()
    if chars[index] == self.content {
        stop_cut = index;
        continue;
    } else { break; }
}
let new_token: String = chars[start_cut..stop_cut].iter().collect();

self.stop is a raw config value. A token that is entirely content and shorter than stop keeps decrementing index past 0:

• debug: panicked at strip.rs: attempt to subtract with overflow
• release: index out of bounds: the len is 1 but the index is 18446744073709551615

Separately, when the leading and trailing windows overlap on a short token, start_cut can exceed stop_cut, so the final chars[start_cut..stop_cut] panics on a reversed slice range (slice index starts at 2 but ends at 1).

Both are reachable through Tokenizer::decode for any tokenizer whose decoder is a Strip with stop >= 1. Decoders aren't wrapped in catch_unwind, so one such token crashes the whole decode/decode_batch.

Fix

Bound the trailing loop with self.stop.min(chars.len()) and clamp the range with stop_cut.max(start_cut) so an over-stripped token collapses to an empty string. Existing valid inputs are unaffected (the current tests still pass).

Tests

Added regression tests for the short all-content token (Strip::new('_', 0, 2) on ["_"], and a mixed batch) and the overlapping start/stop windows (Strip::new('H', 2, 1) on ["HH"]). Both panic on the pre-fix code and pass with the fix.

cc @ArthurZucker @Narsil