test: per-tensor {1,0} BFP wire-allocator branch test + Softmax BFP-wire inference pin

Follow-up 5 from PR #413 (BFP epic #410) — the two final-review minors.

1. The four wire allocators (`initLayerOutputs` / `initGradTensor` / `initBufferOutput` / `initBufferInput`) gained BFP arms in PR #413; the per-tensor `{1,0}` sentinel branch (template `groupSize` 0 → whole tensor as one block) has no dedicated test. Add one.
2. Softmax on a BFP wire in the inference path deserves an explicit pin (the storage-dtype guard added in PR #413 covers Softmax backward; the inference-path behavior is currently unpinned).

Start by reading the four allocator entry points—initLayerOutputs, initGradTensor, initBufferOutput, and initBufferInput—and locate the existing BFP wire tests. Add coverage for the per-tensor {1,0} sentinel, then inspect the Softmax inference path and pin its behavior with a BFP wire. Done means both cases have explicit passing tests.

Written by the indexing model from the issue text.