Release Tokenizer 3.6.4

[vthorsteinsson]

Summary

• avoid the redundant second rough-token whitespace pass during normal tokenization
• retain the pass when HTML escape replacement can introduce whitespace
• bump the package version to 3.6.4 and document the performance improvement

Performance

On test/toktest_large.txt (149,040 tokens), median CPython tokenization time decreased from 1.905 s to 1.514 s (20.5%). Profiled calls decreased from 9.32M to 7.77M.

Verification

• pytest: 151 passed
• ruff check .: passed
• mypy src/tokenizer: passed
• source distribution and wheel built successfully
• twine check: passed
• local clean-venv wheel smoke test: passed
• all deep-token fields, including origin spans, matched master on the large corpus