fix: prevent KV cache corruption on SWA/ISWA models + hot-path perf

[avion23]

Problem

Gemma-4 and any model with interleaved sliding window attention (ISWA) crashes on the second call to create_chat_completion:

llm = Llama(model_path="gemma4-q4.gguf", n_gpu_layers=-1)

llm.create_chat_completion(messages=[{"role": "user", "content": "What is 2+2?"}])
# → OK

llm.create_chat_completion(messages=[{"role": "user", "content": "Write a hello world"}])
# → RuntimeError: error during generation: [end of text]

Root cause

ISWA KV caches store position tracking in global maps (g_iswa_pos_max / g_iswa_pos_min) that are cleared by llama_memory_clear() but not by llama_memory_seq_rm(). The generate() method detects a prefix match between consecutive prompts (shared BOS token), calls kv_cache_seq_rm() to remove the divergent tail, sees it return True, and skips the full reset. Stale position maps cause batch allocator inconsistency → llama_decode returns -1.

Additionally, reset() was a no-op on KV cache state (only reset n_tokens), so calling llm.reset() between prompts had no effect.

A separate tokens[:-1] off-by-one in prefix matching silently broke prefix detection for all models.

Fix

1. reset() calls llama_memory_clear() unconditionally — proper KV cache state reset
2. generate() sets longest_prefix = 0 for recurrent/SWA models when the prefix doesn't cover all cached tokens, falling through to full reset
3. Fixed tokens[:-1] → tokens in prefix matching
4. eval() simplified: removed dead code paths, direct 2D logits assignment

Performance

Location	Before	After
set_batch()	Python loop: 512 tokens × 5 assignments	5 numpy bulk writes
_create_completion detokenization	O(n²) per-token detokenize(completion_tokens)	O(1) incremental token_to_piece()
_create_completion logit_bias	np.copy(scores) every token	in-place modification
_create_completion top-k logprobs	sorted() on 128K vocab: O(V log V)	np.argpartition: O(V)
eval() logits	flattened copy + reshape	direct 2D view assignment
token_to_piece()	create_string_buffer(32) with trailing garbage	returns exact byte length

Benchmarks (TinyLlama-1.1B, M1 Pro, 10-50 runs)

Test	Result
Generation throughput	~930 tokens/sec
set_batch (200 token prompt)	~61ms
logit_bias processing	~54ms
top-k argpartition vs sorted (128K vocab)	32x faster (888ms vs 28.3s for 1000 runs)

Testing

• ISWA: Gemma-4-E2B-it (Q4_K_M, Q8_0) — multi-turn 6 rounds ✅
• Ministral-3: 3B — multi-turn 3+ rounds ✅; 8B, 14B — multi-turn 3+ rounds ✅ (requires n_ctx=2048)
• Other: Mistral-7B, Llama-3-8B, Phi-4, Gemma-2-9B, Qwen3-VL-8B, Mistral-Small-24B, Nomos-1, TinyLlama-1.1B ✅ (13 models total)

Functional verification: ISWA crash fix, reset(), prefix matching, logit_bias, set_batch, eval(), token_to_piece, argpartition correctness — all pass.