Fixed unfused attn2d scale

[laclouis5]

This PR fixes the scale parameter of the Attention2d unfused implementation (#2385).

Two aspects were corrected:

• The scale was changed to q.size(-1) ** -0.5, i.e the rsqrt of the number of queries / sequence length. This matches the default value of torch.nn.functional.scaled_dot_product_attention.
• Some transpositions were made in the wrong order, resulting in different results. The implementation now closely matches the PyTorch implementation as described in torch.nn.functional.scaled_dot_product_attention.

Tests were added to check the new implementation against the fused one.

[HuggingFaceDocBuilderDev]

The docs for this PR live here. All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update.