Adds Triton and Flex Attention backward pass implementations

Extends backward equivalence testing to include Triton and Flex Attention implementations alongside existing Python and CUDA versions.

Updates function signatures to return attention bias gradients and removes softmax log-sum-exp calculations for consistency across implementations.

Fixes attention bias application in Python reference implementation and improves gradient retention handling for proper backward pass computation.

Enhances test configurations with comprehensive parameter combinations and better error handling for missing implementations.

Author: LoserCheems