Add fused MRotaryEmbedding contrib op for Qwen mRoPE variants

[Copilot]

Description

Introduces a new com.microsoft contrib op, MRotaryEmbedding, that fuses the multimodal RoPE (mRoPE) subgraphs used by the Qwen vision-language model family into a single op, replacing the multi-node Shape/Gather/Unsqueeze/Expand/MatMul/Concat/Cos/Sin/Split/Gather/Slice/Neg/Concat chain currently built per Q/K tensor by the onnxruntime-genai model builder.

Schema (bert_defs.cc, ms_opset.h)

• Generalizes RotaryEmbedding: position_ids becomes a required 3D (3, batch_size, sequence_length) tensor stacking Temporal/Height/Width position streams.
• New required attr mrope_section (3 ints, sum == rotary_embedding_dim/2).
• New optional attr mrope_layout (0 = Sectioned/Chunked — Qwen2-VL/2.5-VL, 1 = Interleaved — Qwen3-VL/3.5 family).
• Retains scale, interleaved, rotary_embedding_dim, num_heads, is_packed_batching from RotaryEmbedding so a single full-width section reduces to standard RoPE.

CPU kernel (contrib_ops/cpu/bert/mrotary_embedding*)

• ComputeDimAssignments precomputes, per cos/sin column, which T/H/W stream owns it (based on mrope_section/mrope_layout), independent of runtime data.
• Per-token: gathers position id for all 3 streams, builds combined cos/sin buffers by column selection, applies scale, then rotates via MlasRotaryEmbedOneRow.

CUDA kernel (contrib_ops/cuda/bert/mrotary_embedding*)

• Per-thread stream selection (contiguous-range for Sectioned, modulo-3 for Interleaved) avoids needing a lookup table on GPU.
• Mirrors the existing RotaryEmbeddingBSNH launcher (grid/block/shared-memory-for-in-place strategy).

Both kernels are registered for float/MLFloat16 (CPU) and float/MLFloat16/BFloat16 (CUDA).

Motivation and Context

Qwen2-VL, Qwen2.5-VL, Qwen3-VL(-MoE), and Qwen3.5(-MoE) each require a distinct, hand-built mRoPE subgraph (~15-20 ops per Q/K rotary application) in the ONNX graph produced by the model builder. Fusing this into one op removes that per-layer overhead, reduces graph size, and gives a single, versioned contract for all current and future mRoPE layout variants.

[azure-pipelines]

Azure Pipelines:
There may be pipelines that require an authorized user to comment /azp run to run.

[github-actions]

You can commit the suggested changes from lintrunner.

[github-actions]

Suggested change

	"(3, batch_size, sequence_length), got ", position_ids_dims.size(), " dimensions",
	"(3, batch_size, sequence_length), got ",
	position_ids_dims.size(), " dimensions",

[github-actions]

Suggested change

	int mrope_section[3]; // section sizes for T, H, W (sum == rotary_embedding_dim / 2)
	MRopeLayout mrope_layout; // how the 3 sections are combined
	int mrope_section[3]; // section sizes for T, H, W (sum == rotary_embedding_dim / 2)
	MRopeLayout mrope_layout; // how the 3 sections are combined