feat(gallery): add Qwen3.8 Cyber GGUF variants

[localai-org-maint-bot]

Description

Add Qwen3.8-27B Uncensored Cyber with two llama.cpp builds: cyjin-yl's IQ4_XS quantization and philbert440's Q8_0 release. Group them through variants and include the matching BF16 vision projector in each entry. Use the embedded Jinja template, publisher sampling defaults, and a 32K default context. Document installation and explicit variant selection.

Notes for Reviewers

• Original model, publisher GGUFs, and IQ4_XS quantization.
• Pin IQ4_XS to d82fb040934e0a491600a49477114d005accf59f, and Q8_0/projector to b2dc3ee9a0216898684d8d0607009626662724d4. All three SHA256 values match both HF LFS metadata and pinned-download x-linked-etag headers.
• IQ4_XS is requantized from Q8_0 using an importance matrix calibrated on coding-agent conversations. Use the ordinary file; no grafted MTP weights or speculative-decoding claims.
• Inspected IQ4_XS GGUF metadata: general.architecture=qwen35, native context 262144, embedded chat template. LocalAI's pinned llama.cpp has the corresponding loader and graph. No backend changes.
• Focused gallery variant invariants pass. YAML validation confirms all 1,890 existing entries remain unchanged and verifies file paths, checksums, pinned revisions, and variant references.
• Full go test ./core/gallery -count=1: 392 passed, 3 failed, 9 skipped. Unchanged master reproduces the same three external GitHub/gist fixture failures (Forbidden).
• git diff --check passes. Model inference was not run.

Signed commits

• Yes, I signed my commits. Human DCO attestation is required; the commit includes Assisted-by under repository policy.
• Documentation updated (docs/content/) for user-facing changes, or not applicable