2.0.0-beta-4

• BREAKING CHANGE: Removed experimental deserializer interface and
  supporting code. Applications using this interface should migrate to the
  Unmarshaler interface by implementing UnmarshalMaxMindDB(d *Decoder) error
  instead.
• Open and FromBytes now accept options.
• BREAKING CHANGE: IncludeNetworksWithoutData and IncludeAliasedNetworks
  now return a NetworksOption rather than being one themselves. These must now
  be called as functions: Networks(IncludeAliasedNetworks()) instead of
  Networks(IncludeAliasedNetworks). This was done to improve the documentation
  organization.
• Added Unmarshaler interface to allow custom decoding implementations for
  performance-critical applications. Types implementing
  UnmarshalMaxMindDB(d *Decoder) error will automatically use custom decoding
  logic instead of reflection, following the same pattern as
  json.Unmarshaler.
• Added public Decoder type and Kind constants in mmdbdata package for
  manual decoding. Decoder provides methods like ReadMap(), ReadSlice(),
  ReadString(), ReadUInt32(), PeekKind(), etc. Kind type includes
  helper methods String(), IsContainer(), and IsScalar() for type
  introspection. The main maxminddb package re-exports these types for
  backward compatibility. NewDecoder() supports an options pattern for
  future extensibility.
• Enhanced UnmarshalMaxMindDB to work with nested struct fields, slice
  elements, and map values. The custom unmarshaler is now called recursively
  for any type that implements the Unmarshaler interface, similar to
  encoding/json.
• Improved error messages to include byte offset information and, for the
  reflection-based API, path information for nested structures using JSON
  Pointer format. For example, errors may now show "at offset 1234, path
  /city/names/en" or "at offset 1234, path /list/0/name" instead of just the
  underlying error message.
• PERFORMANCE: Added string interning optimization that reduces allocations
  while maintaining thread safety. Reduces allocation count from 33 to 10 per
  operation in downstream libraries. Uses a fixed 512-entry cache with per-entry
  mutexes for bounded memory usage (~8KB) while minimizing lock contention.