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 paths; wrong scaling or kernel routing would silently change training numerics across models.

Overview
Adds configurable attention scaling (AttentionConfig.scaling / softmax_scale) and threads it through flash varlen, padded flash, SDPA (scale=), ring FA3/FA4, and Ulysses so non-default scales (e.g. upcoming DeepSeek V3) stay consistent across kernels and CP patches.

Refactors flash/SDPA into FlashAttentionCore / SDPAAttentionCore (kernel + _attention_core) with FlashAttention / SDPAAttention owning Q/K/V projections—same hook point for ring/Ulysses _compute_attention overrides on model-specific classes.

Behavior change: when cu_seqlens is absent, flash now uses standard flash_attn_func on [bs, sl, nh, dim] (batch > 1 without CP); varlen remains bs == 1 only. Sliding window is driven via window_size_left on config/core.

Adds CUDA unit tests comparing FA2/3/4, ring varlen, Ulysses varlen, and sliding-window cores against SDPA (default and custom scale).

^{Reviewed by Cursor Bugbot for commit c15bc1d. 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.