tp: shrink the executor's setup and per-batch code

The executor carried several kilobytes of code which buys nothing:
setup which runs once per query, per-batch bookkeeping, and plan-time
casts, all compiled at -O3 with their helpers inlined into every
caller. None of these changes touches a loop over rows.

- Optionals in operator state are plain values: a vector of optionals
  cannot be reset with a memset, so -O3 unrolled each reset into
  kilobytes of code. Key types and column shapes are learnt together
  from the first batch, so an empty vector says they are unknown.
- Selection's and RowBatch's copying moves out of line, as each of a
  dozen operators carried its own vectorized copy, and index copies
  become memcpy and memmove.
- A pipeline segment is a range of steps rather than two vectors, and
  states are built at their size.
- Filter value casters are one per storage type rather than one per
  type and comparison, 36 functions down to 6.
- AssertType builds its errors out of line, and ColumnView::At checks
  for a sequence only when reading uint32_t, as only Id columns are
  sequences.