Fix a wrong reference to seqlen_k variable in the fwd_splitkv kernel

[muoshuosha]

Use params.seqlen_k is wrong with varlen cases，this bug had been fixed in vllm-project/flash-attntion

[muoshuosha]

params.seqlen_k also is wrong in page attn when use cu_seqlen_k store the real seqlen_k

flash-attention/csrc/flash_attn/flash_api.cpp

Lines 1386 to 1393 in 6b1d059

	if (seqlens_k_.has_value()) {
	auto seqlens_k = seqlens_k_.value();
	TORCH_CHECK(seqlens_k.dtype() == torch::kInt32, "seqlens_k must have dtype int32");
	CHECK_DEVICE(seqlens_k);
	CHECK_CONTIGUOUS(seqlens_k);
	CHECK_SHAPE(seqlens_k, batch_size);
	params.cu_seqlens_k = static_cast<int *>(seqlens_k.data_ptr());
	}