Add decision support (decide/choose/score) for Granite Switch models #1741

Summary

Granite Switch models ship a decision adapter: a trained capability that answers fixed-answer questions (is this urgent? which team handles this? how severe, 0-2?) with a single token, read from the model's probability distribution over the answer options. One forward pass per question, fast enough to use inside control flow.

Mellea should expose this as plain Python.

Proposed API

from mellea.stdlib.components.intrinsic import decisions

# Yes/no - a probability you can threshold:
if decisions.decide(ticket, backend, "Does this require urgent attention?") >= 0.7:
    escalate(ticket)

# Multiple choice - best option plus the full distribution:
route = decisions.choose(ticket, backend, "Which team handles this?",
                         {"payments": "Payout failures and refunds",
                          "account": "Login and account access"})

# Rating - expected level plus per-level probabilities:
sev = decisions.score(ticket, backend, "How urgent is this?",
                      ["can wait a week", "this week", "today"])

state (the first argument) is the thing being judged: a string, a dict or list, or a whole conversation. Answers are typed, and a distribution that doesn't concentrate on the declared options raises instead of returning a plausible-looking wrong number.

Context

• The decision adapter ships in the Granite Switch 4.1 checkpoints (announcement blog post)
• Serving requires vLLM with --max-logprobs 20 and the granite-switch package; also works with llama.cpp's llama-server
• The readout reads the model's top-20 token logprobs at the answer position and renormalizes over the option labels

Out of scope

• Accuracy/calibration benchmarking (model-team domain; needs a labeled evaluation set)
• A default revision pin table for Hub checkpoints