[pkubaj]

The big-endian branches were never compiled and no longer build:
direct_bitset.rs uses copy_nonoverlapping without importing it and
gates lvl_get_item to little-endian while its callers are unconditional,
bitset/serialization.rs calls a non-existent chunked_read with a
non-existent data_blocks_len. The remaining branches are
unimplemented!().

Implement them so that the serialized format is the same on both byte
orders (every u64 lane stored little-endian, lanes in array order):

• BitBlock::to_le_bytes/from_le_bytes for the wide SIMD blocks convert
  lane-wise. BitBlock::to_le now reinterprets the little-endian bytes
  natively (from_ne_bytes) instead of round-tripping back to the
  original value; BufferedWriter relies on that.
• Writer::write_primitives/write_masks and
  Reader::read_primitives/read_masks/foreach_read_mask convert per
  element on big-endian. foreach_read_mask needs T: BitBlock for that.
• BitSet::deserialize reads the data blocks through foreach_read_mask on
  big-endian.
• DirectBitset::read_mask reads the byte array and converts it. The old
  ALIGNED branch also passed size_of::() as a count of masks, not
  bytes. lvl_get_item needs no index correction: sub-masks are stored as
  little-endian u64 in array order and read_mask:: already converts
  them.

Tested on FreeBSD/powerpc64 (big-endian, POWER9): the library tests
pass, and data serialized on x86-64 deserializes identically through
BitSet, ImmutableBitset and DirectBitset (aligned and unaligned), and
vice versa.

[tower120]

Do you REALLY need big-endian? How do you test it?