Update llama.cpp to v0.2.0 and modernize the Rust API

[mcharytoniuk]

Summary

• pin the official llama.cpp v0.2.0 release and bump the workspace to 0.13.0
• replace deprecated and legacy entry points with one error-aware Rust API
• map wrapper contract failures, native statuses, allocation failures, and exceptions into typed Rust errors
• simplify native source discovery, CMake configuration, linking, and rebuild tracking across Linux, macOS, Windows, and Android
• reorganize integration tests by behavior and exclude only the pure integration-test crate from coverage accounting

Validation

• make coverage TEST_DEVICE=cuda RUST_TEST_THREADS=1
• all 13 LLM model phases passed serially on CUDA
• every configured coverage gate passed; introduced production Rust executable lines are 1489/1489 covered
• cargo clippy --workspace --all-targets --features cuda -- -D warnings
• cargo fmt --all -- --check
• strict clang-tidy and cppcheck validation passed for project C++ wrappers
• packaging validation passed for publishable leaf crates

GitHub workflows and the Makefile are unchanged. CI runs build and unit validation only; it does not run LLM tests.