Add torch scaled dot product attention (FlashAttention)

[Copilot]

Adds support for PyTorch's scaled dot product attention (FlashAttention) to the torch backend, along with single-line formatting improvements across returnn/torch/frontend/_backend.py.

Changes

• New Function: Added scaled_dot_product_attention method to TorchBackend class that utilizes torch.nn.functional.scaled_dot_product_attention for efficient attention computation
• Frontend Integration: Added corresponding scaled_dot_product_attention function in returnn/frontend/_backend.py and returnn/frontend/attention.py
• Formatting: Applied single-line formatting style for function signatures, list comprehensions, and method calls across the torch backend (152 insertions, 477 deletions)

The new attention implementation includes:

• Support for attention masks and causal masking
• Configurable dropout and scaling
• Fallback to legacy implementation for edge cases (e.g., dynamic dimensions, debugging mode)
• Proper dimension handling for batch, spatial, and embedding dimensions

This enables more efficient attention operations by leveraging PyTorch's optimized FlashAttention implementation when available.

---

💡 You can make Copilot smarter by setting up custom instructions, customizing its development environment and configuring Model Context Protocol (MCP) servers. Learn more Copilot coding agent tips in the docs.

[dorian-K]

@copilot You reverted formatting changes only of the last commit e5f636c, but you were supposed to revert the formatting only changes that were made between e5f636c and 884edfe. The latest commit of my pr already reverted some previously made changes, so your revert of my revert did entirely the wrong thing

[Copilot]

@copilot You reverted formatting changes only of the last commit e5f636c, but you were supposed to revert the formatting only changes that were made between e5f636c...

You're right, I misunderstood the request. I've now restored the file to the e5f636c state, which includes the single-line formatting AND the new scaled_dot_product_attention function. The changes are in commit c1b58ca.

[dorian-K]

Your pull request is now empty, you didnt change anything?

[dorian-K]

@copilot Your pull request is now empty, you didnt change anything?

[Copilot]

@copilot Your pull request is now empty, you didnt change anything?

You're correct - my commits resulted in no net changes because the final state matches e5f636c, which appears to be set as the PR base. The changes from e5f636c (the new scaled_dot_product_attention function and formatting improvements) should be part of the PR content, not the base.

Could you clarify what the correct base commit should be for this PR? Should it be comparing against master branch (commit 884edfe), or is there a different base I should use?

[albertz]

I like the good work here. Well done.