[npu] 43 weight matrices quantized to int8 and resident on the NPU

--- greedy story (CPU fp32 stream, both backends fed identically) ---
One day, a robot named Bolt found a small blue key in the garden. He picked it up and looked around. He saw a big tree with a hole in it. He thought, "Maybe this key will fit in the hole."
He put the key in the hole and turned it. The hole opened and he saw a big, shiny box. He was very happy. He said, "Wow! I found a box with a lock. I want to see what is inside."
He opened the box and saw a lot of toys. He was very happy. He said, "This is my box. I found it in the garden. I want to play with it."
He took the box and ran to his house. He opened the box and saw many things. He saw a ball, a doll, a car, a book, and a hat. He said, "This is my box. I found it in the garden. I want to play with it."
He put the box on the grass and ran to his house.

[fidelity] 201 decode steps
  raw-logit cosine       mean 0.992944  min 0.299905 (step 32, top-1 agreed there)
  mean-centered cosine   mean 0.997243  min 0.885507
  softmax KL(cpu||npu)   mean 0.00792  max 0.34282 nats
  top-5 set overlap      4.80/5  (identical top-5 in 80% of steps)
  argmax disagreements: 5/201 (2.5%), first at step 65
    step  65: cpu 'W' vs npu 'I' — cpu pick is rank 1 on NPU, fp32 logit gap +0.0128
    step  86: cpu ' opened' vs npu ' tried' — cpu pick is rank 1 on NPU, fp32 logit gap +0.2139
    step 188: cpu ' put' vs npu ' took' — cpu pick is rank 1 on NPU, fp32 logit gap +0.2338
    step 193: cpu ' grass' vs npu ' ground' — cpu pick is rank 1 on NPU, fp32 logit gap +0.0191
    step 195: cpu ' ran' vs npu ' played' — cpu pick is rank 1 on NPU, fp32 logit gap +0.1791
  (disagreements land on coin-flip tokens — the fp32 top-2 gap is within int8 noise there, not a broken matmul)
  note: per-channel int8 dequant shifts the whole logit vector by a small DC offset;
        over 32000 vocab entries that alone can drag raw cosine to ~0.3 while the
        top-5 and the softmax are untouched. Center before you trust a cosine.
