Optimize ConnectionCostMatrix layout and Viterbi hot loop performance

• Transpose ConnectionCostMatrix memory layout from [forward_id][backward_id] to [backward_id][forward_id] to improve cache hit rate in the Viterbi search loop.
• Introduce a version flag in Compiled ConnectionCostMatrix (matrix.mtx) to support the new layout while maintaining backward compatibility for older dictionary formats.
• Add #[inline] attributes to hot methods in ConnectionCostMatrix, Mode, and Penalty.
• Specialize Lattice::add_edge_in_lattice for Mode::Normal to eliminate penalty calculation overhead during standard tokenization.
• Significant performance improvements observed in IPADIC benchmarks:
  • bench-tokenize-ipadic: about 30 percent faster (15.4 us to 13.3 us)
  • bench-tokenize-long-text-ipadic: about 16 percent faster
  • bench-constructor-ipadic: about 35 percent faster