Merge branch 'ggml-org:master' into master

docs: add dotsocr support communication drafts

add dots.ocr (DotsOCRForCausalLM) GGUF conversion and inference support

Add full support for rednote-hilab/dots.ocr, a Qwen2 text backbone (1.7B)
with a modified Qwen2-VL vision encoder (1.2B, 42 layers).

Key architectural differences from Qwen2-VL:
Text model uses standard Qwen2 with 1D RoPE (not M-RoPE
Vision encoder uses RMSNorm (not LayerNorm), SiLU gated MLP,
  Conv2D patch embedding, no attention bias, no window attention
2D M-RoPE internal to vision only with sections [d/4, d/4, 0, 0]
Post-trunk RMSNorm before merger (unique to dots.ocr)
Distinct chat template: <|user|>...<|endofuser|> and
also Address PR Feedback