Add SmolLM2-135M and bloomz-560m tokenizer.json fixtures

Byte-identical tokenizer.json from HuggingFaceTB/SmolLM2-135M@93efa2f097d58c2a74874c7e644dbc9b0cee75a2 and bigscience/bloomz-560m@a2845d7e13dd12efae154a9f1c63fcc2e0cc4b05 (main as of 2026-11-06). Both are byte-level BPE vocabularies with no token for some bytes valid UTF-8 can contain (SmolLM2: 0x04 0x06 0x13 0x14 0x16 0x1D 0xF1 0xF2; bloomz: 0x0B). The tk-convert oracle in huggingface/tokenizers (PR #2550) checks they encode like the released crate, which drops those bytes.

fixtures/models/bloomz-560m.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:3fa39cd4b1500feb205bcce3b9703a4373414cafe4970e0657b413f7ddd2a9d3
+size 14500438

fixtures/models/smollm2-135m.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:9ca9acddb6525a194ec8ac7a87f24fbba7232a9a15ffa1af0c1224fcd888e47c
+size 2104556