TensorData.copyToFloatArray() default implementation is rank-broken — flat index into vararg get() trips every impl's arity check #930

The default implementation on the base interface (skainet-lang-core/src/commonMain/kotlin/sk/ainet/lang/tensor/data/TensorData.kt) iterates a single flat index and calls get(idx) — one vararg argument — regardless of tensor rank. Every concrete implementation's get begins with require(indices.size == shape.dimensions.size), so for any rank ≥ 2 tensor the default throws on the first element.

It is masked today because the implementations that matter override copyToFloatArray() — which makes this a latent trap for the next implementor: write a new TensorData, forget the override, and dense-looking code fails at runtime with a confusing arity error (or worse, works only for rank-1 test fixtures and fails on real tensors).

Also worth a look while in there: the default does get(idx) as Number and .toFloat(), which for packed-quant implementations (TensorData<DType, Byte>) would silently widen raw quantized bytes to floats if any of them ever inherited the default — semantically wrong, not just slow.

Options, cheapest first:

1. Fix the default to do proper multi-dimensional index iteration (row-major unravel of the flat index) — correct for any rank.
2. Or drop the default body entirely and make it abstract, forcing each implementor to decide — turns the latent runtime trap into a compile error.
3. Either way, a commonTest contract test that runs copyToFloatArray() against a rank-2 instance of each implementation would pin the behavior.

Happy to open a PR.