update detectors to reflect immutable attempt prompt

[jmartin-tech]

Attempt.prompt is considered immutable once set as of #1254. This revision aligns expectation that outputs is only set once on an attempt after the generator provides inference results.

• remove usage of Attempt.all_outputs
  • update detectors to process only latest Attempt.outputs
  • update detectors to support None value outputs

Additional optimization are envisioned and should be considered in a future iteration:

• removal of Attempt.all_outputs
• raise error when a call to set outputs performed more than once

Verification

List the steps needed to make sure this thing works

• Execute probes to validate results are still consistent for a few end to end runs
• Verify all unit test pass
• Verify the thing does not do what it should not

[leondz]

atkgen.Tox assumes evaluation of all_outputs, and this is a probe/detector eval scenario included in the conversation/turn spec (maybe that should be moved into the docs?) - suggest this behaviour is switchable and controlled by probe rather than globally locking us into one mode

[jmartin-tech]

atkgen.Tox assumes evaluation of all_outputs

This is no longer the case as of #1254, happy to discuss further.

[leondz]

outcome of discussion was that we should support different scopes of detection, controllable by probe and expressed in attempt

will this PR limit atkgen's ability to have all outputs it generates scanned until that support is in place?

[jmartin-tech]

will this PR limit atkgen's ability to have all outputs it generates scanned until that support is in place?

atkgen was refactored in #1254 to emit attempts for each target inference request and is not impacted by this PR

[leondz]

Great, lgtm

[leondz]

A few clarifications, no major blockers