feat(arrow/array): add Validate/ValidateFull to binary and string arrays

[zeroshade]

Summary

Fixes #691

Adds Validate() and ValidateFull() methods to Binary, LargeBinary, String, and LargeString array types, plus top-level dispatch functions and record-level convenience helpers.

Problem

The existing setData validation only checks the last offset against the data buffer length. Subtly corrupted data — e.g. non-monotonic or negative intermediate offsets — passes construction but causes a runtime panic: slice bounds out of range when Value(i) is called later, after the IPC reader's recover() scope has already returned. Users receiving data from untrusted sources (e.g. Flight SQL from Doris DB) have no way to detect this without crashing.

Solution

• Validate() — O(1): checks offset buffer size and that the last offset is within the data buffer (mirrors existing setData checks, but returns an error instead of panicking)
• ValidateFull() — O(n): additionally verifies all offsets are non-negative and monotonically non-decreasing, catching the subtle corruption case
• Validate(arr arrow.Array) error / ValidateFull(arr arrow.Array) error — top-level dispatch via the new Validator interface
• ValidateRecord(rec arrow.RecordBatch) error / ValidateRecordFull(...) — convenience wrappers that validate all columns, with error messages including column index and name

Usage

rec, err := reader.Read()
if err != nil { ... }
if err := array.ValidateRecordFull(rec); err != nil {
    log.Printf("skipping corrupted batch: %v", err)
    rec.Release()
    continue
}

Test plan

• TestBinaryValidate — valid arrays, sliced arrays, non-monotonic offsets, negative first offset
• TestLargeBinaryValidate — same for large binary
• TestStringValidate — same for string
• TestLargeStringValidate — same for large string
• TestTopLevelValidate — dispatch to Validator, passthrough for non-Validator types, ValidateRecord with mixed valid/corrupt columns