Reconsider the use of md5() and uniqid()

[cmb69]

This has been triggered by the discussion regarding the vote to deprecate md5(), sha1() and uniqid().

We're using md5() and uniqid() for caching purposes, what might cause collisions, so perhaps increasing the entropy would be a good idea. To avoid overly long filenames, we might use base32 encoding (base64 encoding isn't suitable for case insensitive file systems), instead of base16.