bfp test debt: backward staging-width fixtures + permutation-aware reduce fixtures

Follow-up from BFP epic PR5 (#435, review ledger "known deferrals" 3 + task-1 deferral). Two suite blind spots, both "tests missing, code believed correct" — one test-debt issue.

## 1. Epic-wide: no backward test exercises FLOAT32-operand staging WIDTHS

When a FLOAT32-stored operand enters the funnel under `ARITH_BFP`, it is staged (quantized) at the op's anchor widths (outputQ forward / propLossQ backward — the R-N1 rule). The width effect of that staging is pinned exactly once, in a *forward* LayerNorm fixture. A regression in *backward* staging (e.g. anchor mix-up propLossQ↔outputQ, or width fallback to a default) would keep the whole suite green.

Scope: targeted fixtures in the GEMM-family and norm backward tests whose gold values are sensitive to the staged width (two widths whose results differ, per the forward LN-B precedent).

## 2. Reduce BFP arms: no permutation-aware fixture

`meanOverTrailingAxesBfp` / `varianceBiasedOverTrailingAxesBfp` are tested only with identity `orderOfDimensions`. BFP blocking binds to storage order, and the zero-copy-transpose hazard is THE structural BFP trap — yet no reduce test exercises a transposed view.

Scope: reduce fixtures on permuted shapes (storage order != logical order) with golds derived from the storage-order walk.

Part of the #410 epic hygiene trail.

🤖 Generated with [Claude Code](https://claude.com/claude-code)

Start by reading the existing forward LayerNorm fixture, then inspect the GEMM-family and norm backward tests and the meanOverTrailingAxesBfp and varianceBiasedOverTrailingAxesBfp reduce tests. Add backward fixtures covering two staging widths and permuted-shape reduce fixtures with storage-order golds; the suite should exercise both blind spots.

Written by the indexing model from the issue text.