[hakarrr]

Fixes #1229.

Send User-Agent: llama-cpp-python when fetching remote images in Llava15ChatHandler and the independent MTMDChatHandler loader. The Llava-derived handlers inherit the same change. This identifies the package instead of relying on urllib's default Python-urllib/... user agent; it does not impersonate a browser or claim to fix every server's HTTP 403 policy.

The change only wraps the URL in urllib.request.Request before opening it. Data-URI decoding is unchanged. File URLs, response bytes and HTTP errors retain their existing behavior. Redirect handling stays with urllib; the added regression checks HTTP 302, not every redirect status.

Validation

• New tests use a loopback HTTP server that rejects requests without the expected package user agent, including a redirect to the image endpoint. No external image service is called.
• Both independent loaders are checked for exact returned bytes, user agent across redirects, HTTP 404 propagation, unchanged data-URI decoding without network access, and file-URL loading.
• New tests against the unmodified source: 4 failed, 6 passed. The failures are the direct HTTP and redirected fetches for each loader.
• Full patched python -m pytest tests -q: 36 passed, no skips (26 existing tests plus 10 new cases), including the existing native/model-backed suite.
• Fetched all three published files byte-for-byte, confirmed the exact reviewed patch, then reran the full 36-test suite and lint/format from that branch.
• ruff check llama_cpp tests, ruff format --check llama_cpp tests, and git diff --check: passed.

Environment: Python 3.11.17, Linux x86_64 CPU, GCC 11.4, CMake 3.22.1; pinned llama.cpp 7d7f0287530bc8981690f361d8c1627dbae1c99a, built with GGML_NATIVE=OFF, GGML_OPENMP=OFF, GGML_BLAS=OFF. No new dependency or native-code change. The added tests exercise image fetching, not multimodal image decoding or vision inference.