2.6.0

• Fixed a denial-of-service issue where a crafted database could use repeated
  pointers to cause excessive CPU and memory use during reflection decoding.
  The decoder now limits decoding work and decoded payload size.
• Made search-tree verification visit shared subtrees only once, bounding work
  while still rejecting cycles and overlong paths.
• Added the maxsize:N struct-tag option to limit maps, arrays, strings, and
  bytes in reflection and generated decoders, plus bounded cursor reads.
  Field names containing commas must now be single-quoted.
• Made generated decoders reject duplicate recognized map keys.
• Added mmdbdata.Cursor.Offset() to retrieve a value's resolved control-byte
  offset for caching within a database. It returns an error when resolution
  fails; the legacy Decoder.Offset() retains its original-offset fallback.
• Improved performance:
  • Reduced 28-bit search-tree lookup overhead with single-word node reads.
  • Extended bounded cursor string fast paths to wider data pointers.
  • Avoided repeated reflection dispatch for pointer-backed strings.
  • Inlined compact header reads when skipping values during budgeted decoding.
  • Avoided repeated type dispatch for unsigned integers decoded into any.
  • Cached validated struct-field matches to speed up repeated decoding.