codec: add pre-encoded voice reference (--ref-spk / --ref-rvq)

qwen-codec --talker extracts the speaker embedding (.spk, raw f32)
and the ICL codes (.rvq) in one pass, encode truncated to the hop
boundary conforming to the --ref-wav path. qwen-tts loads them via
--ref-spk / --ref-rvq and skips the speaker encoder and codec encode
on every synthesis: TTFA 205 ms -> 89 ms. Extends qt_tts_params with
ABI v2 latent fields, adds qt_num_codebooks(), ships freeman.spk +
freeman.rvq and switches clone scripts to the latent path. Output is
bit-identical to the raw path at fixed seed.

README.md

@@ -82,6 +82,22 @@ Voice cloning (`clone.sh`, Base, reference WAV plus its transcript) :
     --lang English -o out.wav < prompt.txt
 ```
 
+Pre-encoded reference (`clone.sh`): `qwen-codec --talker` encodes a reference
+WAV into two compact latents in one pass, the `.spk` speaker embedding and
+the `.rvq` ICL codes, bit-identical to what the `--ref-wav` path computes
+internally. Passing them via `--ref-spk` / `--ref-rvq` skips the speaker
+encoder and the codec encode on every synthesis:
+
+```
+build/qwen-codec --model models/qwen-tokenizer-12hz-Q8_0.gguf \
+    --talker models/qwen-talker-1.7b-base-Q8_0.gguf -i ref.wav
+build/qwen-tts \
+    --model models/qwen-talker-1.7b-base-Q8_0.gguf \
+    --codec models/qwen-tokenizer-12hz-Q8_0.gguf \
+    --ref-spk ref.spk --ref-rvq ref.rvq --ref-text ref.txt \
+    --lang English -o out.wav < prompt.txt
+```
+
 Named speaker (`customvoice.sh`, CustomVoice) :
 
 ```