block_quant: quantize einsum weights at either input slot

[czoli1976]

The block_quant transform assumed the constant weight was always EinSum input 0 (tract's canonical gmk,nk form), so on imported ONNX activation @ weight matmuls — where the weight is input 1 — it wired the block-quant tensor and the injected group axis to the wrong operands and tripped the EinSum input-rank check. This made block_quant unusable on essentially any imported transformer (a BERT/MiniLM encoder fails on the first attention matmul). The slot-0 assumption went unnoticed because the only coverage, the matmul_q40 suite, constructs its weights at slot 0.

Fix: derive the weight and activation slots from the matched input and wire them accordingly. Adds three regression tests covering both ONNX orientations (mk,kn, batched bmk,kn) and the canonical mk,nk, each asserting numerical equivalence to X @ Q4_0(W).

🍍