[Kernel] Classification cleanup: unify _jit_ naming, drop empty/model groups, add elementwise (RFC #29630) (#32148)

Co-authored-by: Claude Opus 4.8 <noreply@anthropic.com>

benchmark/kernels/bench_fused_gate_sigmoid_mul_add.py

@@ -7,7 +7,7 @@ over the Qwen3.5 MoE target hidden size.
 import torch
 import triton
 
-from sglang.kernels.ops.layernorm.elementwise import fused_gate_sigmoid_mul_add
+from sglang.kernels.ops.elementwise.elementwise import fused_gate_sigmoid_mul_add
 
 HIDDEN_DIMS = [4096]
 

benchmark/kernels/bench_fused_sigmoid_mul.py

@@ -10,7 +10,7 @@ a fair comparison — the reshape/contiguous cost is included.
 import torch
 import triton
 
-from sglang.kernels.ops.layernorm.elementwise import fused_sigmoid_mul
+from sglang.kernels.ops.elementwise.elementwise import fused_sigmoid_mul
 
 NUM_HEADS = 32
 HEAD_DIM = 256