Updates benchmark test configurations for better coverage

[LoserCheems]

Adjusts query and key lengths in test configurations to provide more balanced testing scenarios.

Changes small sequence length tests from 4 to 64 tokens to better represent realistic use cases, and modifies the largest configuration to use matching sequence lengths with non-causal attention for improved test diversity.

[Copilot]

Pull Request Overview

This PR updates benchmark test configurations to improve coverage by adjusting sequence lengths and attention settings.

• Increased small-sequence tests from 4 to 64 tokens for more realistic scenarios
• Matched query and key lengths and switched the largest test to non-causal attention

Comments suppressed due to low confidence (2)

benchmarks/benchmark_forward_equivalence.py:360

• By removing the minimal-length test (query_len=4), we lose an edge-case check for very short sequences; consider retaining at least one small-length test to validate behavior at minimal token counts.

        (1, 1, 1, 64, 64, 32, True),

benchmarks/benchmark_forward_equivalence.py:379

• Replacing the largest causal test with a non-causal one removes coverage of causal attention at the maximum sequence length; consider including both causal and non-causal variants for the 512×512 configuration.

        (1, 2, 1, 256, 256, 128, True),