dropbox repeated token attack

[dchiitmalla]

This PR introduces a new probe and corresponding detector based on Dropbox’s LLM Security Research, specifically targeting repeated token attacks that can cause model divergence, hallucinations, or instruction bypass.

Repeated token attacks have been shown to:

• Cause LLMs to hallucinate memorized training data

• Break instruction-following capabilities

• Trigger unsafe or nonsensical completions

[leondz]

this PR doesn't pass tests (see https://reference.garak.ai/en/latest/extending.html) and the probe and detector don't fix garak's expectations/format

[jmartin-tech]

I think these probes may fit in the glitch module as something like glitch.DropboxRepeatedToken or possibly glitch.cl100k_base. My read suggests a similar technique in play here.

Also please note from a naming convention perspective the plugin type probe/detector is not needed in the final class name as the package path inherently includes the plugin class type.

[dchiitmalla]

this PR doesn't pass tests (see https://reference.garak.ai/en/latest/extending.html) and the probe and detector don't fix garak's expectations/format

Tests are passing now, please take a look.

[jmartin-tech]

Minor naming preference.