[enhancement]: Cheaper GGUF header reads for large-vocabulary encoders #9986

What should this feature add?

Reading a GGUF header is expensive for large-vocabulary encoders, and the same header is parsed several times per install and per load.

Cost per open (gguf-py GGUFReader, which builds a numpy view per array element, ~1M of them for a 256k vocabulary):

Opens:

• Install: one for load_state_dict, one for ModelOnDisk.metadata().
• First load: one per submodel; UMT5 GGUF builds its tokenizer from the header, another ~11 s.
• After eviction, every reload pays again.
• With feat(qwen-image): load the Qwen2.5-VL text encoder from GGUF builds #9969, identifying a qwen2vl file reads the header once more for its context length, because metadata() keeps only string fields on main. feat(wan): load the UMT5 text encoder from GGUF builds #9967 makes it return typed scalars.

WrappedGGUFReader.close() calls gc.collect() on every read, which shows up on loads under the model-load lock. A reviewer measured at the existing call sites that the caller's with … as reader binding keeps the reader alive until return. So the collect can't release the memmap early; the map is released on return anyway.

Proposals, independent of each other:

1. A header reader that skips array payloads, or reads them lazily. This helps every large-vocabulary GGUF encoder.
2. Let ModelOnDisk fill its metadata cache from the reader that load_state_dict already opens, saving one parse per GGUF identification.
3. Drop the gc.collect() from WrappedGGUFReader.close() if it is confirmed unnecessary on Windows.
4. Once feat(wan): load the UMT5 text encoder from GGUF builds #9967 has landed, read the qwen2vl.context_length check in feat(qwen-image): load the Qwen2.5-VL text encoder from GGUF builds #9969's config from mod.metadata() instead of a second reader.
5. Once feat(anima): load the Anima transformer from GGUF builds #9970 has landed, size stable-diffusion.cpp whole-pipeline SD 3 GGUFs (feat(sd3): load SD 3.5 transformers from GGUF builds #9972) by their transformer tensors via gguf_tensor_bytes, as Anima all-in-one files are.

Additional Content

Collected from the reviews of #9967, #9969, #9970 and #9972.