IQ4_XS: Sharp v22.5.0 template + Q3_K MTP head (was Q8_0)

Body tensors are byte-identical to the previous upload — strip and graft both copy tensor bytes verbatim. Only blk.40 changes, from Q8_0 (0.90 GB) to the same Q3_K-expert head the CyberTiel ladder ships (0.39 GB), plus the chat template moving to qwen3.8-froggeric-v22.5.0. Head precision was swept Q8_0 down to Q2_K with draft acceptance flat to Q3_K.

Tiel-Coder-35B-A3B-MTP-UD-IQ4_XS.gguf

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:bf12bfacb04f587be6eecd578a5dc3d06861797d3a4cf81b1fad3d72b29dbbce
-size 18629540384
+oid sha256:96b4c5430b5a693447a0aafce58840466829e1788e46fc037b23dfbafab009bf
+size 18119933792