Update PREDICATES

[LoserCheems]

This pull request refines the handling of causal masks and predicates in the compute_attn_1rowblock function within the flash_attention_fwd_kernel.h file. The changes improve code clarity, maintainability, and correctness by introducing more descriptive comments, ensuring consistent handling of optional causal masks, and streamlining predicate allocation.

Improvements to causal mask handling:

• Simplified the creation of causal_mask_smem_ptr by reformatting the ternary operation for better readability. ([csrc/src/flash_attention_fwd_kernel.hL238-R240](https://github.com/flash-algo/flash-sparse-attention/pull/8/files#diff-f28f785166139b70460fd44ef7644554c0b332c25122b50ef571322572e86237L238-R240))
• Updated the identity tensor for causal masks (cCausalMask) to use dummy 1×1 dimensions when no causal mask is provided, ensuring consistent tensor allocation. ([csrc/src/flash_attention_fwd_kernel.hL334-R356](https://github.com/flash-algo/flash-sparse-attention/pull/8/files#diff-f28f785166139b70460fd44ef7644554c0b332c25122b50ef571322572e86237L334-R356))

Predicate allocation and comments:

• Added detailed comments to describe the purpose of identity tensors and predicates, such as cQ, cKV, and cZeroHold, improving code readability. ([csrc/src/flash_attention_fwd_kernel.hL334-R356](https://github.com/flash-algo/flash-sparse-attention/pull/8/files#diff-f28f785166139b70460fd44ef7644554c0b332c25122b50ef571322572e86237L334-R356))
• Ensured predicates like tCausalMaskpCausalMask are always allocated, even when not used, to simplify the logic and avoid conditional tensor creation. ([csrc/src/flash_attention_fwd_kernel.hL334-R356](https://github.com/flash-algo/flash-sparse-attention/pull/8/files#diff-f28f785166139b70460fd44ef7644554c0b332c25122b50ef571322572e86237L334-R356))