feat(parquet): skip level materialization for entirely-null columns

When every element in a list, struct, or fixed-size list array is null,
short-circuit level building before the row loop and store a compact
`(def_value, rep_value, count)` tuple on `ArrayLevels` instead of
materializing `Vec<i16>` buffers. The same fast path applies at the leaf
level in `write_levels()` when `logical_nulls` covers every row.

On the write side, `ArrowColumnWriter` detects the `uniform_levels`
tuple and calls a dedicated `write_uniform_null_batch()` that encodes
def/rep levels via `RleEncoder::put_n()` in O(1) amortized time,
bypassing the normal mini-batch chunking and per-element iteration. A
new `LevelEncoder::put_n_with_observer()` fuses encoding with histogram
and counting updates in a single call. `write_uniform_null_batch` chunks
at the configured page row count limit to respect page boundaries.

Also defers `non_null_indices.reserve()` to branches that actually
populate it, avoiding an unnecessary allocation for all-null arrays.

Benchmark results (vs previous commit):

  primitive_all_null/default                  192 µs    (was  8.8 ms, −97.8%)
  primitive_all_null/parquet_2                193 µs    (was  8.8 ms, −97.8%)
  primitive_all_null/zstd_parquet_2           250 µs    (was  8.9 ms, −97.2%)

Signed-off-by: Hippolyte Barraud <hippolyte.barraud@datadoghq.com>

Author: HippoBaro