krea2: don't hardcode the NVIDIA-only cuDNN SDPA backend

[DasPauluteli]

krea2: don't hardcode the NVIDIA-only cuDNN SDPA backend

Problem

extensions_built_in/diffusion_models/krea2/src/mmdit.py forces the cuDNN
attention backend for every attention call:

with sdpa_kernel(SDPBackend.CUDNN_ATTENTION):
    x = F.scaled_dot_product_attention(q, k, v, attn_mask=mask, scale=scale, enable_gqa=gqa)

cuDNN is NVIDIA-only. On any non-NVIDIA backend (AMD ROCm, Intel XPU, Apple MPS)
there is no cuDNN kernel, so every forward pass dies immediately with:

RuntimeError: No available kernel. Aborting execution.
  File ".../krea2/src/mmdit.py", line 60, in attention
    x = F.scaled_dot_product_attention(...)

The model loads, the VAE and Qwen3-VL text encoder run, latents/embeds cache —
and then training/inference cannot take a single step. Krea 2 is unusable on
these backends today.

Fix

Pass a priority list to sdpa_kernel instead of a single hardcoded backend:

with sdpa_kernel([
    SDPBackend.CUDNN_ATTENTION,
    SDPBackend.FLASH_ATTENTION,
    SDPBackend.EFFICIENT_ATTENTION,
    SDPBackend.MATH,
]):
    x = F.scaled_dot_product_attention(...)

• NVIDIA: behaviour is unchanged — cuDNN is still first and gets selected.
• Everything else: the dispatcher skips the unavailable cuDNN kernel and
  falls back to flash / mem-efficient / math.

One-line semantic change; no config/API surface change.

Testing

Verified Krea 2 RAW LoRA training end-to-end on an AMD Radeon 8060S
(gfx1151 / Strix Halo), ROCm 7.2, torch 2.12.1+rocm7.2:

• Before: crashes on the first step with No available kernel.
• After: all training steps run and a valid LoRA .safetensors is written
  (diffusion_model.*.lora_A/lora_B, loads in ComfyUI).

On this GPU the fallback resolves to AOTriton flash for the unmasked image
blocks (with TORCH_ROCM_AOTRITON_ENABLE_EXPERIMENTAL=1) and math for the
masked text-fusion path. NVIDIA users are unaffected (still cuDNN).

Note for other AMD/ROCm users

This PR only fixes the hard crash. To actually run Krea 2 on ROCm you also need
TORCH_ROCM_AOTRITON_ENABLE_EXPERIMENTAL=1 (for the flash path) and to have the
usual audio/video imports (torchaudio, av, librosa, mutagen) installed.
Full recipe + numbers in the linked writeup.