Added softmax_scale in attentions

[Sirorezka]

1. Added softmax_scale in attentions (+ring, ulysses)
2. Refactored Attention function to split Q,K,V calculations and core attention call.
3. Added possibility to use flash attentions for cases when batch>1 (non cp)
4. Added tests on flash attention layer

This fix is required to later merge DeepSeek_V3 (#2681)

---

Note

High Risk
Touches the core transformer attention and distributed ring/Ulysses kernels; incorrect scaling or path selection would silently change model numerics during training.

Overview
Extends attention so custom softmax_scale (via AttentionConfig.scaling) flows through Flash Attention (FA2/3/4), SDPA, ring, and Ulysses varlen paths instead of always using head_dim**-0.5.

Refactors FlashAttention / SDPAAttention into core modules (kernel only) plus wrapper modules that own Q/K/V projections, so ring/Ulysses can keep patching _compute_attention while shared logic lives in one place.

Flash path behavior: when cu_seqlens is omitted, attention uses the dense flash_attn_func path for batched inputs (batch > 1); varlen still requires batch == 1. Adds window_size_left on config for sliding-window flash kernels.

Adds CUDA unit tests comparing FA variants, ring, and Ulysses varlen outputs against SDPA (including custom scale and sliding window).

^{Reviewed by Cursor Bugbot for commit 0b83f36. Bugbot is set up for automated code reviews on this repo. Configure here.}

[cursor]

Cursor Bugbot has reviewed your changes and found 1 potential issue.

^{❌ Bugbot Autofix is OFF. To automatically fix reported issues with cloud agents, enable autofix in the Cursor dashboard.}

^{Reviewed by Cursor Bugbot for commit 3429290. Configure here.}

[Sirorezka]

I appreciate you're very busy, but I would be exceedingly grateful for your feedback on this pull request whenever you get a chance. Thank you! @samsja @mikasenghaas @S1ro1.