2.3.0

• This module now targets Go 1.25+.
• Reduced reflection decoding time and heap allocations on the hot path. A
  city-lookup benchmark decoding a geoip2-style result runs about 15% faster
  and allocates about 39% fewer bytes per lookup compared to 2.2.0.
• Specialized the IPv4 search-tree walk for 24-, 28-, and 32-bit record
  sizes to skip the IPv6 prefix when looking up IPv4 addresses.
• Decoding into a non-nil slice with sufficient capacity now reuses the
  caller's backing array instead of allocating a fresh slice, matching
  encoding/json semantics. Callers that share slice headers across
  Decode calls should be aware that the backing memory is now mutated.
• Reduced contention under concurrent lookups by switching the internal string
  cache to a lock-free design.