Replace Bark Q8 with quality-safe hybrid quantization

Keep semantic and coarse autoregressive transformers plus codec in F16; quantize only fine non-causal dense matrices.

Bark-Small-GGUF/bark-small-q8_0.gguf

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:ad9c85cff5a28a8710c192f60ac5b63abed4a371e003f1cf7e5f0a93a0eff757
-size 571511328
+oid sha256:83d5463ca1e8f7645a253b1ed1507051c8c3fcc4243e96454004a06b2d5555a9
+size 737103168