Collect reproducible memory and retrieval benchmarks

[initial-d]

me_fasttext is now packaged with a benchmark protocol in docs/benchmark_protocol.md. The next useful step is to collect reproducible reports across corpora, languages, and hardware.

Useful reports should include:

• corpus manifest: language, document count, token count, preprocessing, train/eval split
• hardware manifest: CPU, memory, storage, OS, compiler
• build and training commands
• baselines used, ideally original FastText and me_fasttext
• memory, artifact size, cold load time, warm query latency, and task quality

The most useful comparison is not a single speed number. It is whether exact trie-based subword ids plus compact mmap serving reduce deployment cost while preserving OOV behavior and downstream quality.

Please use the benchmark issue template when sharing results.