Use Arrow IO and memory API for better performance

[marcin-krystianc]

• Replace C++ standard library IO (fstream, ofstream, fopen/fread, ostringstream) and memory (std::vector<uint8_t>) with Arrow equivalents (arrow::io::ReadableFile, arrow::io::FileOutputStream, arrow::io::BufferOutputStream, arrow::ResizableBuffer)
• Since arrow uses the mimalloc allocator, we can benefit from its arena allocator (memory reuse)
• Eliminates platform-dependent pubsetbuf workarounds and reduces unnecessary buffer copies (e.g., ostringstream::str() + memcpy into vector<char> is now a single BufferOutputStream::Finish())
• Update Cython bindings to pass arrow::Buffer directly instead of std::vector<char> + memoryview

[adamreeve]

Looks good thanks Marcin, just noticed a minor typo