fix: UTF-8 char boundary panic in repetition detector

[LopezNuance]

Summary

The repetition detector splits the output buffer into two halves using byte arithmetic (len/2, len-600) which panics when the offset lands inside a multi-byte UTF-8 character (math symbols like ᵢ, ×, σ, CJK). This poisons the Mutex<LlamaContext> and every subsequent request to that Shimmy instance returns 502 forever until restart.

This PR also includes the sliding-window KV eviction, char-trigram cosine repetition detection, model caching, and penalty forwarding that were pending in prior PRs on this branch.

Fix

Add floor_char_boundary() — walks back at most 3 bytes to a valid UTF-8 char boundary. O(1), matches the not-yet-stable str::floor_char_boundary() API. Applied at the two panicking slice sites in the repetition detection window splitting.

Note on prior PRs

PRs #187, #192, #193, and #194 all addressed aspects of the same UTF-8 handling issue. #187 fixed from_utf8 in token emission. #192–#194 attempted to fix the repetition detector but used char_indices() on an already-invalid slice (which itself panics). This PR replaces that with the correct approach. Sorry for the noise — squashed to a single signed commit to keep the history clean.

Test plan

• cargo fmt -- --check passes
• cargo clippy --no-default-features --features huggingface,llama -- -D warnings — zero warnings
• cargo test --lib --features llama — 334 passed
• cargo test --lib --features llama ppt — 15 PPT contracts passed
• cargo test --test regression --features llama — 123 passed (1 pre-existing: issue_129 GPU workflow)
• Live smoke test on 2× A6000 (CUDA build): gemma3:1b and cogito:3b pass math-symbol prompts without panic or 502

[LopezNuance]

Closing in favor of a comprehensive fix PR that addresses the root causes of degenerate output: penalty parameter order swap, sampler chain ordering, KV eviction destroying system prompt, and more. Sorry for the noise — the previous PRs were iterating on symptoms while the root causes were being diagnosed.