Remove redundant inner scope so that SPDA count once

[parthash0804]

Purpose

The attention profiling scopes added in #996 cause the ViT TORCH_SDPA backend to report exactly 2x the number of attention calls (and ~2x "CUDA total") compared to the real number of ops. FLASH_ATTN and TRITON_ATTN report correct counts.

Example (Qwen/Qwen3-VL-4B-Instruct, identical config, only --mm-encoder-attn-backend differs, ViT shape H=16 D=64 KV_H=16):

Backend	Calls / shape
FLASH_ATTN	24
TRITON_ATTN	24
TORCH_SDPA	48 ← wrong

Root cause

Every ViT backend is wrapped once at the encoder layer in MMEncoderAttention._forward_{sdpa,fa,triton} via create_attention_profiler_scope(...).

For FLASH_ATTN and TRITON_ATTN, the inner scope lives inside the direct_register_custom_op op body (flash_attn_maxseqlen_wrapper, triton_attn_wrapper), so key_averages() records one event per op.

For TORCH_SDPA, the inner scope was placed in the apply_sdpa helper, which is a plain function called by the torch_sdpa_wrapper custom op. Its record_function is recorded as a separate nested event with the same scope name as the encoder-layer scope (for ViT self-attention q_len == kv_len, so B/Q/S/H/D/KV_H all match). key_averages() aggregates both same-named events → every SDPA op is counted twice, inflating both "Calls" and "CUDA total" by 2x.

This is a measurement/counting artifact only — no extra GPU work is done.

Fix

Remove the redundant profiler scope inside apply_sdpa. The SDPA op is already covered by the _forward_sdpa encoder-layer scope, matching how FLASH_ATTN/TRITON_ATTN are counted. This also fixes per-segment shape mislabeling when apply_sdpa is called once per cu_seqlens chunk.

Test Result

• FLASH_ATTN / TRITON_ATTN: unchanged (24 calls/shape)
• TORCH_SDPA: 48 → 24 calls/shape

[mgehre-amd]

Thanks!