FIX: TF32 warning (#43012)

[shantanugupta2004]

What does this PR do?

This PR replaces the matrix multiplication operator (@) with broadcasting element-wise multiplication (*) in the RotaryEmbedding implementation for several major models (Llama, Mistral, Mixtral, Qwen2, Gemma, Gemma2).
When compiling a model with torch.compile in bfloat16, the RoPE frequency calculation (which is intentionally kept in float32 for precision) triggers a UserWarning regarding TensorFloat32 (TF32) if it's not enabled.
Since the shapes involved in this specific operation [batch, dim/2, 1] and [batch, 1, seq_len] result in an outer product, using @ is mathematically equivalent to * with broadcasting. However, using * avoids the "matrix multiplication" code path in the compiler, effectively silencing the false-positive warning and potentially offering a minor performance optimization by avoiding a full GEMM call for a simple outer product.

Fixes #43012

Before submitting

• This PR fixes a typo or improves the docs (you can dismiss the other checks if that's the case).
• Did you read the contributor guideline,
  Pull Request section?
• Was this discussed/approved via a Github issue link?
• Did you write any new necessary tests?

Who can review?

Anyone in the community is free to review the PR once the tests have passed.

[github-actions]

[For maintainers] Suggested jobs to run (before merge)

run-slow: afmoe, apertus, arcee, aria, bamba, bitnet, chameleon, csm, cwm, dbrx, deepseek_v3, dia, diffllama, doge, dots1, emu3