Document the input chunk alignment for decode_mut

[riha-dev]

decode_len documents chunk-by-chunk decoding but omits the chunk alignment that encode_len documents, so an unpadded base whose bit-width does not divide 8 decodes a trailing partial group of symbols as the end of the input and silently drops the rest (BASE64_NOPAD fed AQ then AA gives [1, 0], where decode(b"AQAA") gives [1, 0, 0]); decode_align now provides that alignment, mirroring encode_align.