Do not panic in the Strip decoder on crafted config or short tokens

[v-code01]

What

Strip's content/start/stop fields are deserialized straight from the config (tokenizer.json) with no validation, and decode_chain trusts them, so a crafted decoder config (plus tokens from the model vocab) panics:

let mut stop_cut = chars.len();
for i in 0..self.stop {
    let index = chars.len() - i - 1;   // underflows when a token is shorter than `stop`
    ...
}
let new_token: String = chars[start_cut..stop_cut].iter().collect(); // panics on a reversed range

• Underflow — chars.len() - i - 1 underflows for a token shorter than stop: a panic in debug/overflow-checked builds, and an out-of-bounds index after wrapping in release. An empty token with stop >= 1 triggers it ("attempt to subtract with overflow").
• Reversed range — when a token is made entirely of content characters, start_cut overruns stop_cut, so chars[start_cut..stop_cut] panics. E.g. Strip::new('H', 2, 2) decoding "HH" → chars[2..0] ("slice index starts at 2 but ends at 0").

Fix

Bound the trailing loop by the token length (0..self.stop.min(chars.len())) and clamp stop_cut = stop_cut.max(start_cut) so the slice range is never reversed. Both crafted cases now decode to an empty string, consistent with "everything was stripped".

Tests

Adds does_not_panic_on_crafted_config covering the reversed-range and empty/short-token cases; the existing decode test is unchanged.

cc @ArthurZucker @McPatate