feat: update llama.cpp to v0.4.0

[abetlen]

Update the vendored llama.cpp to the latest versioned release, v0.4.0 (5266f24da).

• Synchronize model-loading, quantization, and MTMD ctypes layouts and bind the new MTMD input-parts and placeholder APIs.
• Pass the new media-helper options in the example server while retaining default options for existing Python helper calls.
• Inherit upstream MTMD library version variables and update the native state-format constants and changelog.

Previously serialized native context states must be regenerated after this update.

Validation: CPU build, 26 tests, C/ctypes ABI checks, Qwen 0.8B MTMD image smoke test, and Ruff checks passed locally.