model: use F.scaled_dot_product_attention for ViT-family attention

[rebel-junamsong]

Summary

Replace manual matmul → softmax → matmul attention in ViT-family vision modules with F.scaled_dot_product_attention. This routes the op through rbln-compiler's dedicated SDPA custom-op path (mapped to __paged_normal_*_sdpa_{5d,6d}_ kernels) instead of the generic relay fallback, improving numerical stability under compDtype=bfloat.

Affected models

• qwen2_vl — VisionAttention
• qwen2_5_vl — Qwen2_5_VLVisionFullAttention, Qwen2_5_VLVisionWindowAttention
• pixtral — PixtralAttention

Qwen3VLVisionAttention already uses SDPA. GroundingDINO variants use non-standard attention (dual-softmax / grid_sample) and are not applicable.

4D vs 5D tensor rank

Each model keeps its original Q/K/V rank — no forced unification. rbln-compiler dispatches to different kernels based on rank:

Input rank	Internal layout	Kernel
4D [B, H, L, D]	NHCW64c	__paged_normal_*_sdpa_5d_
5D [B, H, 1, L, D]	NDHCW64c	__paged_normal_*_sdpa_6d_

The SDPA converter path in rebel.core.custom_converter._pt_attention has no rank assertion, so both 4D and 5D work.

Model	Rank	Kernel
Qwen2-VL	4D	_5d_
Qwen2.5-VL Full / Window	4D	_5d_
Pixtral	5D	_6d_
Qwen3-VL (unchanged)	4D	_5d_

Test plan

• Qwen2-VL / ColQwen2 retrieval — compile + accuracy
• Qwen2.5-VL image/video — compile + correlation vs native
• Pixtral vision encoder — compile + accuracy
• Qwen3-VL regression (shares compiler path)