feat(amber): columnar wire format, operator contract, scan and filter

Feature Summary

In one sentence: lay the foundation of columnar execution, the Arrow wire format and the opt-in operator contract, and prove it end to end with two operators (the CSV scan and the filter).

Parent: #8556 (opt-in columnar execution). This is PR 1 of the stacked series and pairs with apache/texera#8558.

What is this?

Before any operator can be made faster, two things must exist: a way to carry a batch of columns between workers, and a contract that lets an operator opt in to reading those columns. This issue adds both, then wires up the two simplest operators (scan and filter) so the whole path can be tested honestly.

Think of it as laying one lane of a new highway and driving two cars down it, while the old road stays open and default.

Proposed Solution or Design

The two new primitives.

• ColumnarFrame: a data payload that carries one Arrow batch (the raw column bytes) plus a row count. It rides alongside the existing per-row DataFrame.
• ColumnarOperatorExecutor / ColumnarResult: the opt-in contract. An operator can implement processColumnarBatch(...) and answer Emit (I made a new batch), Consumed (I absorbed it), or Unsupported (decode it and run the row path).

%%{init: {'theme':'dark', 'themeVariables': {'background':'#000000','lineColor':'#0F766E'}}}%%
flowchart LR
  SCAN[CSV scan: read file straight into Arrow columns] --> WIRE[ColumnarFrame on the wire]
  WIRE --> FILT[filter: keep rows on the column, no per-row decode]
  FILT --> TERM[terminal]
  classDef default fill:#000,color:#fff,stroke:#888,stroke-width:1px

How the engine routes a batch. The DataProcessor (the per-worker loop that runs the operator) checks whether the operator understands columns; if not, it decodes to rows. OutputManager, the partitioner, and the input side carry the batch end to end.

%%{init: {'theme':'dark', 'themeVariables': {'background':'#000000','lineColor':'#000000'}}}%%
flowchart TD
  F{ColumnarFrame and<br/>operator opts in?} -->|yes| C[processColumnarBatch]
  F -->|no| R[decode to rows, row path]
  classDef default fill:#000,color:#fff,stroke:#888,stroke-width:1px
  style C stroke:#1B7F3B
  style R stroke:#B0451E

What lands here. ColumnarFrame, the contract, ArrowUtils (serialize/deserialize a batch), engine plumbing (DataProcessor, OutputManager, partitioner, input side), the Arrow-producing CSVScanSourceOpExec, and the Arrow-consuming SpecializedFilterOpExec. All behind a flag; default off.

Verified: the Arrow round-trip is lossless across all types, the vectorized filter matches the row filter exactly, and the existing DataProcessingSpec passes with the flag on and off.

High-level overview. Part of #8556.