perf: add -DNDEBUG compilation flag (#998)

We observed some performance degradation after upgrading cutlass to
v3.9.

This is because some runtime checks are added to cutlass in debug mode:
*
https://github.com/NVIDIA/cutlass/blob/df8a550d3917b0e97f416b2ed8c2d786f7f686a3/include/cutlass/pipeline/sm90_pipeline.hpp#L66-L70
*
https://github.com/NVIDIA/cutlass/blob/df8a550d3917b0e97f416b2ed8c2d786f7f686a3/include/cutlass/pipeline/sm90_pipeline.hpp#L82-L86

This PR addresses the issue by adding `-DNDEBUG` to compilation flags in
both JIT and AOT mode.

After this PR, we no longer observe the performance degradation.

Author: yzh119

flashinfer/jit/core.py

@@ -114,6 +114,9 @@ def load_cuda_ops(
             "--ptxas-options=-v",
             "--ptxas-options=--verbose,--register-usage-level=10,--warn-on-local-memory-usage",
         ]
+    else:
+        # non debug mode
+        cuda_cflags += ["-DNDEBUG"]
 
     cflags += extra_cflags
     cuda_cflags += extra_cuda_cflags

setup.py

@@ -225,6 +225,7 @@ def __init__(self, *args, **kwargs) -> None:
         "-Xfatbin",
         "-compress-all",
         "-use_fast_math",
+        "-DNDEBUG",
         "-DPy_LIMITED_API=0x03080000",
     ]
     libraries = [