Optimize ShapeTesselator speed and memory with C++17 idioms

Replace std::ostringstream with std::to_chars + pre-reserved std::string
in export functions (~37% faster). Use std::memcpy for bulk float copies
in GetVertices/NormalsAsTuple (~23% faster) and JoinPrimitives. Skip
identity transforms in ProcessSingleFace. Store Face/Edge by value
instead of unique_ptr to eliminate per-face heap allocation. Remove
redundant reserve() before resize() calls. Add [[nodiscard]] on getters.
Remove unused includes (memory, tuple, sstream, iomanip, numeric, mutex).

Author: tpaviot