Simplify anonymizer and prepare for Windows support

[woct0rdho]

There are some changes to the anonymizer that I think reasonable:

1. Make anonymize_path an alias of anonymize_text. We don't need to treat paths differently from general text.
2. Even if a special folder like Documents, Downloads, Desktop is in the path, we don't need to remove the parent folders before it. We can keep the absolute path and only redact the username.
3. When len(username) >= 4, we only need to globally replace the username with the hash. This is a superset of all other replacements in anonymize_text. Only when len(username) < 4, we replace with more specific patterns.
4. I've unified the handling of /Users/username (macOS), \Users\username , \\Users\\username (Windows path that may be escaped in various forms), -Users-username (hyphen-encoded path like Claude Code and HuggingFace cache). I've also unified the handling of /Users/username and /home/username.
5. We can supply a custom home dir to anonymize_text. Only when it's not any conventional home dir above, we use a specific pattern to handle it.

The rest is to optimize the speed using re.compile and precomputed regex patterns. The resulting anonymize_text function looks pretty clean. You can see the updated tests for the effect of the anonymizer.

This is manually written by me, not vibe-coded. I've prepared the remaining changes of Windows support (based on this PR) at https://github.com/woct0rdho/dataclaw/tree/windows , and an example dataset at https://huggingface.co/datasets/woctordho/dataclaw-windows . But before that, we should carefully review this PR because it touches the anonymizer.

By default the only scope of the anonymizer is to redact the operation system's username. Other kinds of PII are inherently subjective to define, and the user is responsible for defining strings to be redacted.

Currently all CI tests run on Ubuntu. We may need to let some of them run on macOS and Windows, for example Python 3.10 + Ubuntu, Python 3.11 + macOS, Python 3.12 + Windows.