~6% speedup on decode_slice/3

No perf impact

- Rearrange main decoding loops to handle chunks of 32 bytes at a time, then 4 bytes at a time, meaning that `decode_suffix` need only handle 0-4 bytes, simplifying its code. Moderate speed gains of around 5-10%.
- Improve error precision. `InvalidLength` now has a `usize` length indicating how many valid symbols were found, but that the count of those symbols was invalid. Before, it just did `input % 4 == `, which was harder to reason about, as there might be padding etc. DecoderReader now also precisely reports the suitable InvalidByte if an earlier block of decoding found padding that was valid in that context, but more padding was found later, rendering that earlier padding invalid.
- Tidy up decode tests. There were some duplicated scenarios, and certain aspects are now tested in more detail.

Implementations must now precisely, not conservatively, return an error when the output length is too small.

The logic isn't important, but it helps make actual coverage gaps more visible.