* passing a non-string secret or non-string hash to any
  CryptContext or handler method will now reliably result
  in a TypeError.

  previously, passing hash=None to many handler identify() and verify()
  methods would return False, while others would raise a TypeError.
  other handler methods would alternately throw ValueError or TypeError
  when passed a value that wasn't unicode or bytes.
  the various CryptContext methods also behaved inconsistently,
  depending on the behavior of the underlying handler.

  all of these behaviors are gone, they should all raise the same TypeError.

* redid many of the from_string() methods to verify the hash type.

* moved secret type & size validation to GenericHandler's encrypt/genhash/verify methods.
  this cheaply made the secret validation global to all hashes, and lets
  _calc_digest() implementations trust that the secret is valid.

* updated the CryptContext and handler unittests to verify the above behavior is adhered to.

* support variable salt size of 4-16 bytes [issue 30].
* throw correct error when invalid base64 encoding is encountered.
* added some test vectors for the above.

* studied crypt_blowfish's 8bit bug
	- verified none of passlib's backends were affected
	- added recognition (but not support) for crypt_blowfish's $2x$ hash prefix
	- added support for crypt_blowfish's $2y$ hash prefix
	- note in docs about Passlib's current handling of crypt_blowfish 8bit issues.
* refactored bcrypt's salt-unused-bits repair code into Base64Engine.repair_unused(),
  making the code cleaner and more isolated. a bunch more tests.
* added bcrypt64 (bcrypt-base64 variant) to utils
* added LazyBase64Engine to reduce memory / startup time