[localai-org-maint-bot]

Add cdiamond's Qwen3.8-27B iMatrix NVFP4 GGUF as a variant of qwen3.8-27b-q4. It uses llama.cpp, includes the F16 vision projector, and enables the embedded MTP head with one draft token. The default context is 32,768 tokens. The publisher's repository had about 3.16 million Hugging Face downloads when checked.

Also fix qwen3.8-27b-nvfp4-mtp: its mmproj pointed to a Q8_0 filename, but its files list downloads the BF16 projector. The configuration now uses the downloaded file.

Sources: publisher's model card, original model. Both new downloads are pinned to ac343e8f44caef0896f79d372ecc07ef7ab34ec8; each SHA256 matches both HF API LFS metadata and the pinned URL's x-linked-etag. Checked current master and open gallery PR diffs for duplicates.

Validation:

• YAML parsing, unique names, variant references and ownership passed across 2,209 entries.
• Model and projector file references passed for both affected entries.
• git diff --check passed.
• go test ./core/gallery/... could not start: proxy.golang.org returned HTTP 403 for the LocalAGI module.
• No inference smoke test was run.

Documentation updated in docs/content/features/model-gallery.md.

Human DCO sign-off remains required. The commit uses Assisted-by; repository policy prohibits AI-added Signed-off-by trailers.