fix(io,cpu): infer BPE for legacy tokenizer.json (#858); gather with N-D indices (#859)

[michalharakal]

Fixes the last two bugs surfaced while porting an educational GPT project.

#858 — TokenizerFactory rejects legacy tokenizer.json

fromTokenizerJson required a model.type field and threw UnsupportedTokenizerException otherwise. GPT-2's official openai-community/gpt2 tokenizer.json predates that field, so it couldn't be loaded through the factory (the underlying QwenByteLevelBpeTokenizer.fromTokenizerJson parsed it fine).

Fix: when model.type is absent, infer it from structure. A model.merges list is unique to BPE among the HF model types (Unigram and WordPiece have none), so its presence routes the file to QwenByteLevelBpeTokenizer. Files with neither a type nor merges still throw, with a clearer message.

#859 — CPU gather fails on multi-dimensional indices

The op documents [N, L] index support, but read the indices with a flat indices.data[i] — the vararg element accessor needs one coordinate per dimension, so rank-2 indices threw Number of indices (1) must match tensor dimensions (2).

Fix: read the indices in row-major order — directly from the contiguous IntArray/FloatArray buffer when present, otherwise by unraveling the flat position into a coordinate. The output-shape logic was already correct for rank-2 indices.

Tests

• TokenizerFactoryDispatchTest: a legacy no-model.type tokenizer.json (with a byte-level pretokenizer and a merges list) infers BPE and encodes correctly; a file with neither type nor merges still throws.
• GatherRowDequantTest: gather with [2,3] indices over a [10,4] table returns the expected [2,3,4] result.
• Full jvmTest for skainet-io-core and skainet-backend-cpu green.

With this, all eight issues the port surfaced have a fix (merged or in review).