Should support any buffer objects, not just bytes

Use ffi.from_buffer(...) for conversions, so that bytearray, memoryview and other things work too, instead of checking that the arguments are bytes instances.

I would also prefer being able to give output buffers to low level functions, for instance I had to do this hack to enable encryption without creating a new buffer. It works even in place, if ciphertext and message are the same buffer:

from nacl._sodium import ffi, lib

def encrypt_into(ciphertext, message, aad, nonce, key):
  mlen = len(message)
  clen = ffi.new("unsigned long long *")
  ciphertext = ffi.from_buffer(ciphertext)
  message = ffi.from_buffer(message)
  if aad:
    _aad = ffi.from_buffer(aad)
    aalen = len(aad)
  else:
    _aad = ffi.NULL
    aalen = 0

  return lib.crypto_aead_chacha20poly1305_ietf_encrypt(
    ciphertext, clen, message, mlen, _aad, aalen, ffi.NULL, nonce, key
  )

While for most things it does not matter if copies are made and new buffers are being allocated, stream ciphers can be much faster with proper buffer management.

Start by reviewing the low-level Python binding functions that currently require bytes and compare them with the shown ffi.from_buffer usage. Define which input and output buffer cases must work, including bytearray, memoryview, and in-place encryption, then verify those cases without unnecessary buffer allocation.