Transpose the encoding matrix

Stripe together the encodings from each of the 16 codewords, so that
if a 512-byte DRAM line is knocked out, it affects 256 bits from each
codeword rather than 4096 bits from a single encoded codeword.

Rather than using std::bitset, we directly set and read bits in
the std::vector<uint8_t>, because the striping means that copying it
will now cost not4k in allocation but 64k.

Decode directly to a word, without using list decoding. It seems
we don't need list decoding for the error rates that matter here,
and we never completed the implementation of it anyway.

Declare and test only the full interface, now that it doesn't decompose
quite so neatly.

Bug: 63928581
Test: atest HadamardTest
Change-Id: If022d3f4a8d6fccdf68119d4666f83ce5005bccb