models: model_infer puts a whole raw output tensor in the message — let a call gather/argmax it natively first #370

Summary

A model_infer call with raw: true writes each output tensor into the message whole. When a workflow needs only a few values from it — a per-cell policy head [1,5,H,W] read at the handful of cells a caller cares about, then an argmax — the whole tensor then rides every later step of the pass, and every step pays for it (#371).

Size of it

A [1,5,124,120] f32 head is 297 KB. In a turn loop that asks 8 seats, those heads (plus the copies the workflow makes when it routes them back to their seats) are copied into the evaluation arena on every later async condition, templated field and sync run of the pass: on the order of 60 MB of memcpy per turn before any inference. Unsetting the tensors as soon as they are decoded (which a workflow can do today) removed most of that: a no-op model on the widest board went from 10.6 to 8.7 ms/turn.

Copies along the way, per call: the adapter's tensor to owned; Tensor::from_bytes into tract; the output back to OwnedDataTensor; then every later evaluation copies its bytes again (datavalue OwnedDataTensor::to_arena).

Proposal

Let model_infer reduce an output before it enters the message, natively, with operands templated from the message — for example:

"select": { "output": "policy", "reshape": [5, -1], "gather": { "axis": 1, "indices": { "var": "…" } }, "argmax": 1 }

writing [n] integers instead of the tensor. The existing result expression cannot do this because it sees only the outputs, not the rest of the message (e.g. the indices of the cells to read), which is why callers use raw today.