Fix jagged_hstu_attn mapping and use quick autotuning in TPU benchmark

- Fix wrong kernel function name: jagged_hstu_attn -> _helion_jagged_attention_kernel
- Add HELION_AUTOTUNE_EFFORT=quick to CI workflow — full effort times out
  for 5/11 kernels (welford, attention, geglu, grpo_loss, swiglu)

Author: norx1991

.github/workflows/benchmark_tpu.yml

@@ -17,6 +17,7 @@ jobs:
     env:
       HELION_BACKEND: pallas
       HELION_AUTOTUNE_LOG_LEVEL: INFO
+      HELION_AUTOTUNE_EFFORT: quick
 
     runs-on: linux.google.tpuv7x.1
 

benchmarks/run_tpu.py

@@ -111,7 +111,7 @@ def _softmax_shapes() -> list[tuple[str, tuple[Any, ...]]]:
     "bmm": ("bmm", "bmm", None, None),
     "geglu": ("geglu", "geglu", None, None),
     "grpo_loss": ("grpo_loss", "grpo_loss_forward", None, None),
-    "jagged_hstu_attn": ("jagged_hstu_attn", "jagged_hstu_attn", None, None),
+    "jagged_hstu_attn": ("jagged_hstu_attn", "_helion_jagged_attention_kernel", None, None),
     "low_mem_dropout": ("low_mem_dropout", "low_mem_dropout", None, None),
     "swiglu": ("swiglu", "swiglu_fwd", None, None),
 }