common : store the tokens after each n-gram in a sorted vector

[jadidbourbaki]

Stacked on #5.

I replaced the inner std::unordered_map of the n-gram caches with a vector of (token, count) pairs sorted by token. The inner map sends each token that follows an n-gram to its count. Lookups binary search the vector.

common_ngram_cache_part keeps the find, emplace, iteration, and size calls of common/ngram-cache.cpp, so common/ngram-cache.cpp is unchanged. C++23 has std::flat_map for a sorted vector with a map interface, but llama.cpp builds as C++17.

The drafting logic and the cache file format are unchanged.

Results

The benchmark is running.