Harden rail fence cipher decoding against resource exhaustion

[krotname]

Motivation

• The previous decode implementation performed allocations and O(n^2) work based on the caller-controlled rail count n, and created per-character String/LinkedList objects even for empty input, enabling CPU/memory exhaustion for large n.
• The change aims to prevent resource amplification while preserving the kata's decode semantics for valid inputs (rails >= 2) and typical test cases.

Description

• Validate the rail count by throwing IllegalArgumentException if n < 2, and return immediately when s.isEmpty() or n >= s.length() to avoid unnecessary work and allocations.
• Replace the quadratic sumArr scans and TreeMap/LinkedList per-character conversions with an int[] position offset array and a single StringBuilder result to reconstruct characters directly from the input.
• Remove per-character String allocations and linked-list buffering so decoding runs in linear time relative to the input length plus rail bookkeeping.
• Add a regression test shouldDecodeWithoutAllocatingForUnusedRails asserting that extreme rail counts (e.g. Integer.MAX_VALUE) do not cause allocation or incorrect decoding for empty and short inputs.

Testing

• Ran mvn -Dtest=RailFenceCipherTest test and the RailFenceCipherTest suite completed with all tests passing (5 tests, 0 failures).
• Ran the full test suite with mvn -q test and it completed successfully with no test failures.

---

Codex Task

[krotname]

Superseded by the validated combined merge in #115: #115