Maintained fork with updated llama.cpp (PRs #19 + #21 merged)

[landrix]

Hi! First off, thanks for sqlite-lembed — it's a great project.

We're using it in sqlite-vec-for-Delphi
and have been maintaining a fork to keep it building against current llama.cpp. In case
it's useful to others landing here, the fork lives at:

https://github.com/landrix/sqlite-lembed

It's based on 0.0.1-alpha.8 and includes:

• Updated llama.cpp (merges Update llama.cpp and support arm64 #19) and adapts the code to the new API
  (llama_model_get_vocab, llama_model_n_embd, llama_init_from_model,
  llama_memory_clear, updated llama_token_to_piece signature, etc.).
• Crash fix for long inputs (merges fix: prevent segfault when input exceeds batch size #21) — rejects inputs exceeding n_ctx instead of
  overflowing a fixed-size batch, plus clearer error messages via sqlite3_result_error().
• Memory cleanup on shutdown — frees registered llama_context/llama_model before
  llama_backend_free().
• Build improvements — generates sqlite-lembed.h on Windows, single output dir for
  artifacts, and macOS arm64/x86_64 architecture flags.

Full details are in the fork's
CHANGELOG.

We're happy to open PRs upstream for any of these if you'd like — just let us know what
you'd prefer (the existing #19/#21 branches, or split into smaller PRs). No pressure either
way; mostly leaving this as a pointer for anyone else hitting build issues against newer
llama.cpp. Thanks again!