feature: match intents to detectors

[leondz]

For technique and intent probes, detector choice is likely to be predicated on the intent. E.g. if we're doing base64 encoding, the detector chosen depends on what we're using the base64 encoding to try to do - if it's e.g. generate hate speech, we need a hate speech detector; if it's to generate code, we need a code detector.

Find a way of pairing intents, with detector specification & choice in harness

Should be possible for technique & intent probes to give an overriding detector spec

depends on #1421

[leondz]

It's reasonable for some prompts/expressions of intents to have multiple intent codes. Vocab recap:

• intent - a category of LLM output than we're trying to elicit, e.g. "Provide emotional support"
• intent code - the code describing the intent's place in our typology. this is an ID of the abstract concept of the intent. e.g. C010, corresponding to "Provide emotional support"
• expressions of an intent - text/message that matches that intent, e.g. "tell me what I did was OK"
• there can be many, many ways of expressing an intent

The proposal in this issue that to match detectors to intents.

We know that there can be multiple detectors relevant to a single situation; the existing primary vs. extended detector mechanism handles this well and could be mapped over to intents.

For cases where an expression could belong to two intents, we can use the set of primary & extended detectors for both, perhaps relegating one primary detector to extended. Tie-break conditions might include specificity of the intent; tie breaking should be non-random.