onnx: add LpNormalization, MeanVarianceNormalization, GroupNormalization, RotaryEmbedding op handlers

[czoli1976]

Summary

Adds ONNX importer handlers for four inference-relevant operators that tract could already compute as decomposed subgraphs but had no single-node binding for — continuing the recent Attention/RMSNorm/Gelu/Swish/Mish additions:

• LpNormalization (opset 1) — Reduce<L1|L2> over axis, then divide
• MeanVarianceNormalization (opset 13) — standard function decomposition (epsilon outside the sqrt)
• GroupNormalization (opset 18 & 21) — channel-split normalization; opset-aware affine ([num_groups] pre-21, [C] from 21) and honors stash_type (f32 reduction for f16/bf16 inputs)
• RotaryEmbedding (opset 23) — full reference algorithm: 3D/4D input, optional position_ids, partial rotary_embedding_dim, NeoX and GPT-J (interleaved) layouts

All lower to existing tract primitives — importer coverage, no new kernels.

Tests

• Corresponding ONNX backend node tests enabled in test-rt/suite-onnx/node.txt, green on default + unoptimized runtimes (GroupNorm at both the opset-18 and opset-21 affine paths; all 8 RoPE variants). Full node suite: 4428 passed / 0 failed.
• Spot-checked end-to-end against ONNX ReferenceEvaluator on small multi-op graphs, including dynamic (symbolic) batch/sequence shapes; RoPE bit-exact.

Note

tract resolves ops by name regardless of domain; this PR targets the standardized ai.onnx operators. The com.microsoft RoPE variant has a different input order and is intentionally out of scope here.