Fixes tensor addressing for ZOH and active mask in splitkv

[LoserCheems]

Refactors tensor creation to use pre-calculated offsets instead of local_tile operations, ensuring correct memory addressing for both block table and non-block table scenarios.

Updates pointer advancement logic to maintain proper synchronization between K, ZOH, and active mask tensors during block iteration.

Adds debug output to help diagnose memory addressing issues during development.

[Copilot]

Pull Request Overview

This PR fixes tensor addressing for ZOH and active mask in the splitkv kernel, refactoring tensor creation to use pre-calculated offsets and updating pointer advancement logic.

• Introduces new offset calculations (col_offset_zoh and col_offset_am) to correctly compute tensor pointers.
• Replaces local_tile calls with direct make_tensor constructs and updates pointer arithmetic for ZOH and active mask tensors.
• Adds debug print statements for monitoring tensor pointer values during execution.

[Copilot]

Consider wrapping the debug print statements with a compile-time flag (e.g., #ifdef DEBUG) to avoid unintentional output in production builds.

Suggested change

	);
	);
	#ifdef DEBUG

[Copilot]

Similarly, consider guarding these debug prints with a compile-time condition to prevent extra overhead in production runs.

Suggested change

	binfo.actual_seqlen_k - n_block * kBlockN);
	if (tidx == 0 && bidh == 0 && bidb == 0) {
	printf("Before copy_ZOH: n_block=%d, seqlen_q_offset=%d, seqlen_k_offset=%d\n",
	n_block,
	binfo.actual_seqlen_q - m_block * kBlockM,
	binfo.actual_seqlen_k - n_block * kBlockN);
	}
	binfo.actual_seqlen_k - n_block * kBlockN);
	#ifdef DEBUG
	if (tidx == 0 && bidh == 0 && bidb == 0) {
	printf("Before copy_ZOH: n_block=%d, seqlen_q_offset=%d, seqlen_k_offset=%d\n",
	n_block,
	binfo.actual_seqlen_q - m_block * kBlockM,
	binfo.actual_seqlen_k - n_block * kBlockN);
	}
	#endif