feat(array): scalar probes with one-off and repeated probe types - #9843

Row access over arrays on top of the `probe_scalar` hook:

- `ArrayProbe<'a>` is a bare `&ArrayRef` with no destructor. One-off
  reads, including every child read while recursing through nested
  arrays, own nothing, so there is no unwind cleanup, no out-of-line drop
  glue and no pinned temporary on the per-row path. This is what the
  CodSpeed regression on the earlier design came down to.
- `RepeatedArrayProbe` owns its handle, encoding state, validity probe
  and child probes, created on first use and reused across rows.
- Both implement `Probe`. `ArrayRef::probe()` and
  `ArrayRef::repeated_probe()` are the entry points; `execute_scalar` and
  `is_valid` are shims over `probe()`.
- `ProbeState<'_, V>` carries the `ArrayView` and, for a repeated read, a
  borrow of the retained `RepeatedState`; it has no destructor.
  `state.array()`, `state.retained()`, `state.slot(i) -> impl Probe`, and
  the direct `child_scalar` / `child_is_valid`. Encodings never branch on
  policy for child reads.
- `Validity::probe()` gives the retained validity accessor.
- `Struct` and `Primitive` implement `probe_scalar`; `scalar_at`
  delegates through `ProbeState::once`.

Not generic over the retention policy: a generic vtable method is
instantiated in every crate that constructs arrays of the encoding, so
the hot body compiles downstream without this crate's inlining; measured
at 20-30 points on struct reads and rejected. The policy is a runtime
`Option` check, folded into the one-off arm by `inline(always)` on the
probe bodies.

Local divan medians vs the develop tip, interleaved runs of 1000
samples: is_valid_per_element[1024] -10%, [256] -12%,
execute_scalar_struct_wide -39%, execute_scalar_struct_simple -29%.

Signed-off-by: Joe Isaacs <joe.isaacs@live.co.uk>
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_016CqrLKgPqFYGZK5sjk1qe7

…tate

- `ArrayProbe<'a>` is now an enum: `Once(&ArrayRef)` for a one-off read,
  `Repeated(&mut RepeatedArrayProbe)` for a borrow of a retained probe.
  Both variants are references, so it still has no destructor, which is
  the property the per-row numbers depend on; a compile-time
  `needs_drop` assertion pins that for `ArrayProbe` and `ProbeState`.
- `ProbeState::slot(i) -> VortexResult<Option<ArrayProbe>>` is the child
  accessor: `None` for an absent slot, an error for an out-of-bounds one.
  `ProbeState::split()` returns the encoding state and a `ProbeChildren`
  as disjoint borrows, so a cache can stay borrowed while children are
  read. The `child_scalar` / `child_is_valid` shortcuts are gone; reading
  `Struct` fields through `slot()` measures within noise of them.
- `ProbeState::once` is crate-private: it exists for the `scalar_at`
  delegation and goes away with `scalar_at`.
- The `Probe` trait and the private `ChildProbe` enum are gone; the enum
  is the one borrowed reader and `RepeatedArrayProbe::as_probe` gives the
  retained variant where a probe is expected.
- `ProbeValidity` and `Validity::probe` are gone: `RepeatedArrayProbe`
  resolves validity once into `uniform_validity: Option<bool>` or a
  boxed validity probe.
- `ProbeStorage` is gone; the erased state slot is an
  `Option<Box<dyn Any>>` with a crate-private `repeated_state` accessor.
- `RepeatedState` is crate-private, and the retained side lives in
  `probe/repeated.rs`.
- No `inline(always)` anywhere: the per-row bodies are plain `#[inline]`.
  That leaves the one-off shim one thin frame above the encoding, which
  measures at parity with develop on `is_valid_per_element`; forcing the
  bodies inline was worth about 10% there and is deliberately not taken.

Public probe items: `ArrayProbe`, `RepeatedArrayProbe`, `ProbeState`,
`ProbeChildren`, and the `EncodingProbeState` alias.

Local divan medians vs the develop tip, five interleaved runs of 1000
samples: is_valid_per_element at parity, execute_scalar_struct_wide
about -34%, execute_scalar_struct_simple about -26%.

Signed-off-by: Joe Isaacs <joe.isaacs@live.co.uk>
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_016CqrLKgPqFYGZK5sjk1qe7