talker, code predictor: persistent KV cache

Adds a per-layer K/V ring (kv-cache.h) backed by a dedicated backend
buffer, sized at init for max_seq_len positions. The talker holds a
4096 position cache (896 MB f32) for the LM context, the code
predictor a 16 position cache (~80 KB) reset every frame.

talker_forward splits into prefill (resets the cache and writes T
positions in one shot) and decode (appends one position, reads the
[0, cur_len+1) window). code_predictor_step does the same with a T=2
prefill plus 14 single token decodes.

Bit identical audio output, validated by sha256 against the pre KV
cache run on a 64 frame F32 reference seed=42. Walltime drops ~10%
on a single utterance ; the win scales with sequence length and
unlocks frame by frame streaming.

src/pipeline-tts.h

@@ -11,6 +11,7 @@
 #include "code-predictor-weights.h"
 #include "ggml-backend.h"
 #include "gguf-weights.h"
+#include "kv-cache.h"
 #include "pipeline-codec.h"
 #include "speaker-encoder-weights.h"
 #include "talker-weights.h"
@@ -88,6 +89,12 @@ struct PipelineTTS {
     BackendPair          bp;
     ggml_backend_t       backend;
     ggml_backend_sched_t sched;
+
+    // Persistent KV caches : the talker holds the LM context, the
+    // predictor holds one frame's 16 sub-steps and gets reset every
+    // frame in code_predictor_step.
+    KVCache talker_kv;
+    KVCache code_predictor_kv;
 };
 
 // Open the talker GGUF and the codec GGUF, load every module on the