Speed up lookup by ~13% (template callbacks + root dispatch table)

[yhirose]

Summary

Two format-neutral optimizations to the read path. The compiled byte code is unchanged (byte-identical output; existing .fst files stay fully compatible), so this only touches how a matcher traverses it.

• Template callbacks in matcher::match. The output/prefix callbacks were std::function, so every query constructed a std::function and called through it indirectly. They are now template parameters (with std::nullptr_t defaults and if constexpr guards), so nothing is allocated per query and the calls inline. map::common_prefix_search / set::common_prefix_search are templatized to match (lambda callers are unaffected).
• Root dispatch table. Every query passes through the root state, paying a jump-table binary search on the first character. Each matcher now precomputes a 256-entry label -> address table for the root at construction time, making that step O(1). Costs ~1 KB of memory per matcher; the file is untouched.

Measurements

benchmark/main.cc on /usr/share/dict/words (235,976 keys, Apple M1 Pro), min of 12 runs:

metric	before	after	Δ
exact match (map)	217 ms	189 ms	−12.9%
common prefix (map)	217 ms	189 ms	−12.9%
exact match (keys)	223 ms	196 ms	−12.1%
common prefix (keys)	223 ms	196 ms	−12.1%

Compiled FST size is identical before/after (1,070,382 bytes for map<uint32_t>), confirming the byte code is unchanged.

Testing

• ctest: 36/36 pass.