[snshn]

set_data() takes impl Into<Vec<u8>>, so a Vec<u8> or String gets moved in
instead of copied (slices, arrays and string literals still work).

Serializing writes in chunks instead of building the whole data URL in
memory first; Display uses that, and so does the new write_to(), which
writes to any std::io::Write. encoded_len() gives the exact length of
the serialized data URL without serializing it.

Parsing base64 decodes straight from the input when it has no "%" escapes,
and forgiving-base64 decoding no longer makes a stripped copy of its input.
For a 10 MB payload, peak memory while parsing went from about 40 MB to
10 MB (just the result), and parsing got about a third faster.

The CLI writes its output as it encodes it.