Fix Qwen3 RotaryEmbedding fusion emitting invalid position_ids shape

[Sammy-Dabbas]

The on-the-fly RoPE fusion added for Qwen3 in #27590 wires the traced position_ids into the com.microsoft.RotaryEmbedding node without adapting its shape. A real Qwen3 export has no position_ids input, so HuggingFace builds one with a hardcoded batch axis of 1, shape 1 by sequence_length. The kernel only accepts a scalar or single-element position_ids, or an exact batch_size by sequence_length tensor, so any request with batch_size greater than 1 fails with "Input 'position_ids' dimension 0 should be of size batch_size, got 1". The existing tests missed it because the Qwen3 test generator declared position_ids as an already correct batch_size by sequence_length input.

In on-the-fly RoPE the position_ids is just the contiguous arange, and the kernel's scalar path derives every position from the first element, which reproduces that arange exactly and is batch-independent. This flattens the traced tensor and slices its first element to an equivalent single-offset input, so the fused node is correct at any batch size. A standalone check confirms the reduced output is bit-identical to the intended full arange, and a regression test verifies the fusion now emits it.

Fixes #29618.