feat(sgl-kernel): add InfLLM v2 attention kernels (#29383)

Co-authored-by: Size Wang <paulgeorge13hhhhh@gmail.com>
Co-authored-by: lijiayi <lijiayi@modelbest.cn>
Co-authored-by: suhmily10 <suhmily@gmail.com>
Co-authored-by: Xiaoyue Xu <xiaoyue.xu.me@gmail.com>
Co-authored-by: hansjohn <74091612+hansjohn@users.noreply.github.com>
Co-authored-by: zhangyan <1762895426@qq.com>

sgl-kernel/include/sgl_kernel_ops.h

@@ -112,6 +112,25 @@ int64_t cutlass_mla_get_workspace_size(
     int64_t sm_count = 0,
     int64_t num_kv_splits = 1 /* Set to 1 to avoid cuda_graph issue by default. */);
 
+/*
+ * From csrc/infllm_v2
+ */
+void infllm_v2_max_pooling_1d_varlen(
+    at::Tensor input,
+    at::Tensor output,
+    at::Tensor cu_seqlens_q,
+    at::Tensor cu_seqlens_k,
+    at::Tensor cache_lens,
+    int64_t max_seqlen_q,
+    int64_t max_seqlen_k,
+    int64_t kernel_size,
+    int64_t stride,
+    int64_t padding,
+    int64_t block_size,
+    int64_t local_blocks,
+    int64_t init_blocks,
+    int64_t total_q);
+
 /*
  * From csrc/elementwise
  */