Compute Widen aggregates on the values - #10890

`Mask::last` computed `len - 1` eagerly inside `then_some`, so an empty all-true mask panicked
with a subtraction overflow in debug builds. The `Last` aggregate reaches it for every empty batch
with all-valid validity.

Signed-off-by: Connor Tsui <[email protected]>

Add aggregate kernels for `Widen` that build the exact partial of the
logical dtype from the values, so that `MinMax`, `Min`, `Max`,
`IsConstant`, `IsSorted` and `Sum` with `skip_nans` no longer widen the
values first. The extrema and verdict kernels ask the values and decode
any value they return. The sum kernel sums the values into the sum state
of the logical dtype. It declines values that have a sum kernel of their
own, exact or fallback (`RunEnd`, `Sparse`, `Chunked`), constant values
and batches of more than 2^32 rows. Every other aggregate keeps the
default path.

The kernels are registered in `AggregateFnSession::default` for their
exact aggregate ids. Scalar access and the kernels share one decode
function.

Signed-off-by: Connor Tsui <[email protected]>

Accumulate each aggregate over batches encoded as `Widen` and as
canonical arrays, and check every kernel partial, running partial and
result. Each consumer builds its own batch, so no accumulator reads a
statistic that encoding or another accumulator cached. The cases cover
the child bounds, nullable and non-nullable logical dtypes, batch
boundaries, empty and all-null batches, a single null under strict
`IsSorted`, a running sum overflow, and the sum declining constant and
chunked values. Add a `vortex-fastlanes` integration test with a
`BitPacked` child.

Signed-off-by: Connor Tsui <[email protected]>

Replace the MinMax, IsConstant, IsSorted and Sum kernels of Widen with
one kernel that runs the same aggregate on the values and casts the
partial to the partial dtype of the logical dtype. Decoding is strictly
increasing, one-to-one and keeps validity, so the partials agree, and the
values keep their own cached statistics and aggregate kernels.

Sum is served only when the aggregate has the Stat::Sum slot, so that
each batch is summed from zero, and when the batch has at most 2^32
rows, so that the per-block and per-value overflow checks agree.

Signed-off-by: Connor Tsui <[email protected]>

The aggregate kernel no longer decodes scalars, so scalar access is the only caller. Move the function beside it in the operations vtable.

Signed-off-by: Connor Tsui <[email protected]>

The Widen aggregate kernel runs Sum through its accumulator and no longer builds a sum state itself. Restore SumV2's private sum_state_scalar, the Sum to_scalar match and the doc lines that only the old Widen sum kernel needed.

Signed-off-by: Connor Tsui <[email protected]>

State the Sum overflow bound for 64-bit logical dtypes only, explain that skip_nans matters for integer sums only because it selects the Stat::Sum slot, and say why the chunked values test compares results.

Signed-off-by: Connor Tsui <[email protected]>

Signed-off-by: Connor Tsui <[email protected]>

Signed-off-by: Connor Tsui <[email protected]>

A logical type narrower than 64 bits is summed in the same blocks as its
values, so its overflow checks agree at any length.

Signed-off-by: Connor Tsui <[email protected]>

Sum adds every integer type into a 64-bit sum, so values of b bits cannot
overflow it within 2^(64 - b) rows, whatever the logical width. One rule
replaces the fixed 2^32 bound and the special case for 64-bit logical
types.

Signed-off-by: Connor Tsui <[email protected]>