feat: dynamic sequence length xnnpack export

Replace the static [1,256] xnnpack export with a dynamic-sequence-length one (Dim tokens in [2,256] + get_dynamic_dims_forward companion). ~4.6x faster on short inputs; 100% label agreement vs the PyTorch reference.

xnnpack/privacy_filter_openai_xnnpack_8da4w.pte

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:fcc0bb61dbb074d11bcccb11cc5fcbb624a969945e0a5682e9a25559b137feb2
-size 825207940
+oid sha256:50f4992d3e508d795d2f9645b7d7a8d28d948b0ea821463a7158959fc590dd81
+size 1241705092