Add TensorOps.permute(axes) for arbitrary-axis permutation

[michalharakal]

Context

TensorOps.transpose only swaps the last two dimensions. There's no API today for arbitrary-axis permutation (e.g. swapping dims 0 and 1 of a rank-3 tensor).

This came up concretely in SKaiNET-transformers PR #81: MultiHeadAttention needs to convert [seqLen, nHeads, headDim] → [nHeads, seqLen, headDim] after Q/K/V projection. With no permute available, that PR landed a manual copy-based helper (swapSeqHeadDims) that allocates a fresh FloatArray per call. On a 1024-token prefill of TinyLlama 1.1B, that's ~13M floats / forward (~52 MB) of transient allocation just to fix the shape.

A proper permute(axes) is generic, single-allocation per call, and likely stride-only on backends that support strided tensors.

Proposal

Add to sk.ainet.lang.tensor.ops.TensorOps:

/**
 * Permute the dimensions of [tensor] according to [axes].
 *
 * `axes` is a permutation of `0..tensor.rank-1`; the i-th axis of the
 * result is the `axes[i]`-th axis of the input. For example, on a rank-3
 * tensor of shape `[A, B, C]`, `permute(t, intArrayOf(1, 0, 2))` returns
 * shape `[B, A, C]`.
 *
 * `permute(t, intArrayOf(0, 1, ..., rank-3, rank-1, rank-2))` is
 * equivalent to [transpose].
 *
 * @param tensor input tensor, any rank ≥ 1
 * @param axes a permutation of `0..tensor.rank-1` (length must equal
 *   `tensor.rank`, every value in `[0, rank)` exactly once)
 */
@Diff
public fun <T : DType, V> permute(tensor: Tensor<T, V>, axes: IntArray): Tensor<T, V>

Implementation scope

• commonMain interface declaration + default-shape derivation.
• VoidTensorOps shape-only stub (returns zeros at the permuted shape).
• DefaultCpuOps JVM/native data-moving implementation.
• Unit tests: identity permute, dim-0/dim-1 swap (rank 3), full shuffle (rank 4), round-trip equivalence with permute(permute(t, p), inverse(p)) == t.

Out of scope

• Strided/lazy permute (no data copy). Worth a follow-up once a strided tensor representation lands; for now, copy is fine and matches existing op semantics.
• AD support (the @Diff annotation is included for symmetry with transpose, but the gradient impl can land in a follow-up if needed).

Acceptance

• API compiles on all KMP targets.
• Unit tests pass.
• Downstream MultiHeadAttention.swapSeqHeadDims can be deleted and replaced with ctx.ops.permute(t, intArrayOf(1, 0, 2)).