fix(Tokenizers): avoid quadratic regex matching for added tokens without whitespace padding - #386

Fixes #383.

Summary

In PreTrainedTokenizer, addedTokensRegex was constructed by wrapping every added token in a capturing group (\(token)) even when neither lstrip nor rstrip was requested.

For tokenizers with large added token vocabularies (e.g. Google Gemma-3 with 6,415 added tokens), this caused an $O(N^2)$ quadratic slowdown in Foundation's ICU regular expression engine due to capturing group table allocations, slowing down single newline \n encoding from ~0.02 ms to ~29 ms (~1.1s for a 40-line prompt on device).

Changes

1. Non-capturing groups by default: When $0.prefix and $0.suffix are both false, compile the token using a non-capturing group (?:\(token)) instead of (\(token)).
2. Preserve whitespace stripping: Added tokens with lstrip or rstrip still retain their capturing group so split(by: NSRegularExpression) continues to isolate and strip surrounding whitespace as expected.
3. Empty check guard: Avoid compiling an empty regex pattern when unwrappedAddedTokens is empty.
4. Regression test: Added addedTokensRegexPerformanceAndCorrectness in Tests/TokenizersTests/TokenizerTests.swift.

Verification

• Executed swift test across all targets (50/50 tests passing).
• Validated code style with swift-format lint --strict --recursive Sources Tests (clean, 0 warnings).