Keep i64 off the number paths: lanes fold numbers, cells read as bytes

Counting user-mode instructions per query (bank, 100,000 accounts)
against ee21aec found what i64 had added to the number paths, and an
instruction sampler found where:

- fold_in took a union and asked whether the fold was of i64s for every
  item, in the group and leaf-at-a-time passes too, where a fold of
  i64s never runs (it is never planned). Those passes now call
  fold_number, which takes a double; fold_in keeps the i64 branch for
  folds run item by item. per_branch: 27.83M instructions an ask back
  to 27.05M (26.72M before i64).
- cell_at's third kind made it too big to inline into pure_list_get,
  so every item read became a call. A number's, bool's or integer's
  cell is 8 bytes read as they are (memcpy, so no aliasing rule is
  bent), one branch as before, and it inlines again. A table of 10,000
  rows: 8.78M instructions an ask to 8.70M (8.58M before i64).

Behaviour is unchanged; every check passes.

0 files changed, 0 insertions, 0 deletions