Add Anthropic Claude tokenizer support

[vwilson]

Summary

Closes #32

• Adds a new "claude" encoding based on Anthropic's official @anthropic-ai/tokenizer BPE vocabulary (~65K tokens)
• Adds NFKC text normalization support (the key algorithmic difference between Claude's tokenizer and OpenAI's)
• Adds model name mappings for Claude models (claude-instant-1 through claude-4-sonnet) with prefix matching for dated variants
• Reuses existing Regex50KBase pattern (identical to Claude's pat_str)

Note: Per Anthropic's documentation, this tokenizer is accurate for pre-Claude 3 models and serves as a rough approximation for Claude 3+.

Usage

// By encoding name
var encoding = GptEncoding.GetEncoding("claude");

// By model name
var encoding = GptEncoding.GetEncodingForModel("claude-3.5-sonnet");

var tokens = encoding.Encode("Hello, world!");
var text = encoding.Decode(tokens);

Test plan

• All 889 tests pass (885 existing + 4 new)
• Encode/decode roundtrip works for Claude encoding
• NFKC normalization verified (fullwidth chars normalize correctly)
• Model name resolution works for all Claude variants
• Special token encoding/decoding works
• Zero performance impact on existing encodings (normalization null-check short-circuits)
• Builds on all target frameworks (netstandard2.0, net6.0, net8.0)

🤖 Generated with Claude Code