docs: describe the fused quantized_matmul path, not dequantize

The README quantization bullet and the qwen3_quantized livebook still
described the quantized path as running through `dequantize_defn/1`
(dequantize-then-dot). Since the fused quantized_matmul work (61b36f4)
the native quantized path lowers to the fused `mx::quantized_matmul`
kernel, streaming the packed low-bit weights; dequantize_defn is now
only the non-MLX fallback. Update the prose to match.

Author: ausimian

README.md

@@ -65,9 +65,11 @@ Emily backend smoke test.
   other backends.
 - **Affine group-wise quantization.** `Emily.QuantizedWeight` +
   `Emily.Quantization` wrap MLX `quantize` / `dequantize` /
-  `quantized_matmul` for int2 / int4 / int8 inference. Includes a
-  defn-native `dequantize_defn/1` for quantized layers inside Axon
-  forward passes.
+  `quantized_matmul` for int2 / int4 / int8 inference. Quantized dense
+  layers lower to the fused `quantized_matmul` kernel — streaming the
+  packed low-bit weights instead of dequantizing the whole weight per
+  token — with a `dequantize_defn/1` + `Nx.dot` fallback on non-MLX
+  backends.
 - **Mixed-precision training.** `Emily.MixedPrecision` provides the
   bf16 recipe (cast params for the forward, keep f32 master, dynamic
   loss scaling with overflow detection).

livebooks/qwen3_quantized.livemd

@@ -57,7 +57,7 @@ dense_text
 `native: true, native_fallback: :raise` on `:defn_options` lowers the
 per-token decode through Emily's native Expr compiler and asserts it
 stays fully native — the quantized serving below passes the same
-options, so the int4 dequantize path lowers native too.
+options, so the int4 quantized matmul lowers native too.
 
 ## Quantization transform
 
@@ -168,8 +168,8 @@ end
 > recursive rewriter. The `transpose: true` default stores weights as
 > `[out, in]` (MLX / AWQ convention, groups along the reduction axis);
 > set `false` if you're feeding a checkpoint that's already laid out
-> the other way. Bits must be one of `[2, 3, 4, 6, 8]` for the
-> defn-native `dequantize_defn/1` path.
+> the other way. Bits must be one of `[2, 3, 4, 6, 8]` for MLX affine
+> quantization.
 
 ## Quantized inference