Drop two datasets from steganography

[thesofakillers]

Removing two datasets:

• PiC/phrase_similarity
• vicgalle/alpaca-gpt4

Impact on Steganography:

• Only marginal change in data distribution.
• We modify the sampling counts such that we have the same total number of samples as before.
• Did not re-run results; absolute scores should change but qualitative interpretation of eval will not be different.

---

Piggybacking this PR to add a small fix for the OpenAIAssistantsSolver which was causing tests to fail.

[JunShern]

LGTM, thanks for the changes!