Renames Flash Attention to SDPA in benchmark suite

[LoserCheems]

Updates terminology throughout the performance benchmark to use "SDPA" (Scaled Dot Product Attention) instead of "Flash Attention" for consistency with PyTorch's official naming convention.

Changes function names, variable references, command-line options, and display text to reflect that the baseline implementation uses PyTorch's SDPA rather than specifically Flash Attention, improving clarity about the actual underlying implementation being benchmarked.

[Copilot]

Pull Request Overview

Updates terminology from “Flash Attention” to “SDPA” across the benchmark suite to align with PyTorch’s official naming and improve clarity.

• Renamed function, variables, flags, and display text from Flash Attention to SDPA.
• Adjusted command-line options (--test-type) and titles/headers in output.
• Updated comments and docstrings to reference SDPA baseline.

Comments suppressed due to low confidence (3)

benchmarks/benchmark_forward_performance.py:535

• [nitpick] Consider renaming the variable run_flash to run_sdpa (and updating its usage) to better reflect that it's selecting SDPA runs rather than Flash Attention.

    run_flash = test_type in ['all', 'sdpa', 'sdpa-vs-cuda', 'sdpa-vs-triton', 'sdpa-vs-flex']

benchmarks/benchmark_forward_performance.py:813

• [nitpick] The tuple for 'sdpa' still references 'flash_attention'; consider renaming it (e.g., 'sdpa_attention') or matching the new function name scaled_dot_product_attention_cuda for consistency.

            'sdpa': ('flash_attention', results['flash_attention_status'], results['flash_attention_times']),

benchmarks/benchmark_forward_performance.py:172

• Update the docstring to reflect that this function returns a tuple (attn_outputs, time_ms) or ("OOM", 0) rather than just the attention output, so the documentation matches the actual return values.

    Returns: