feat: support Wan2.2 VACE

[losewayy]

Summary

Issue #1953 asked about Wan2.2 VACE. Testing showed the existing VACE path already handles the new checkpoints correctly once they are loaded — the real obstacles were two problems in the GGUF reading path:

1. Misleading errors when loading >4-dim tensors. Wan2.2 GGUFs (QuantStack) store conv3d weights as 5-dim tensors (patch_embedding.weight, vace_patch_embedding.weight). gguf_init_from_file rejects them with invalid number of dimensions: 5 > 4 and logs an ERROR, even though the extended reader folds them correctly — which made these models look unsupported. Wan2.1_14B_VACE and Wan2.2-T2V GGUFs carry the same 5-dim tensors, so this affects the whole Wan 14B family.

2. A latent bug in the extended reader's ARRAY metadata parsing. It read each array element as a full key-value pair, but GGUF array elements are raw values (no key/type headers). Any file with array-typed metadata (e.g. the ~32k tokenizer vocabulary in umt5-xxl-encoder) fails in that path — which went unnoticed because the extended reader only runs as a fallback.

Changes:

• gguf_reader_ext.h: parse ARRAY metadata by element type (fixed-size seeks / per-string skips); add has_tensors_beyond_ggml_limits().
• gguf_io.cpp: do a cheap header scan first; files declaring >4-dim tensors go straight to the extended reader instead of calling gguf_init_from_file, which is guaranteed to fail. All other files keep the existing behavior (gguf first, reader as fallback).

Also added a docs/wan.md section for Wan2.2 VACE-Fun A14B (MoE expert pair usage) and a generated sample clip.

Related Issue / Discussion

Fix #1953.

Additional Information

Verified locally on an RTX 5070 Ti (CUDA):

• Wan2.2-VACE-Fun-A14B Q3_K_S high/low-noise experts: R2V produces coherent reference-conditioned output (sample in assets/wan/), T2V runs cleanly
• Regular GGUFs (umt5-xxl-encoder Q8_0) load unchanged
• A truncated file still fails cleanly, no crash

Not covered: Vulkan backend (the code path is backend-agnostic; the issue reporter offered to verify), V2V control video (shares the Wan2.1 VACE path).

Checklist

• I have read and confirmed this PR follows the contribution guidelines.

[leejet]

I’ve updated the PR title to use fix:. Wan2.2 VACE is already supported by the existing implementation; this PR fixes GGUF metadata parsing, avoids misleading errors when loading five-dimensional tensors, and adds usage documentation.