[jit_kernel] Migrate copy kernel from sgl-kernel AOT to JIT

[Johnsonms]

Motivation

#17865
copy_to_gpu_no_ce is a small utility kernel that copies a CPU int32 tensor to GPU without using the
copy engine, by packing data into a by-value kernel-argument struct. Migrating it from the AOT
sgl-kernel build to the JIT framework reduces build complexity and aligns with the ongoing effort to
move lightweight kernels to JIT.

Modifications

• Add python/sglang/jit_kernel/csrc/elementwise/copy.cuh: CUDA kernel templated on kN, CPU/CUDA
  tensor validation via TVM FFI TensorMatcher
• Add python/sglang/jit_kernel/copy.py: JIT Python wrapper, module cached per array size via
  @cache_once
• Add python/sglang/jit_kernel/tests/test_copy.py: correctness test for N=64 and N=72
• Add python/sglang/jit_kernel/benchmark/bench_copy.py: latency benchmark
• Update deep_gemm.py: import copy_to_gpu_no_ce from sglang.jit_kernel.copy instead of
  sgl_kernel.elementwise
• Remove sgl-kernel/csrc/elementwise/copy.cu and sgl-kernel/tests/test_copy.py
• Remove AOT registration from CMakeLists.txt, common_extension.cc, sgl_kernel_ops.h,
  sgl_kernel/elementwise.py, and sgl_kernel/init.py

Accuracy Tests

python -m pytest python/sglang/jit_kernel/tests/test_copy.py -v

Benchmarking and Profiling

python python/sglang/jit_kernel/benchmark/bench_copy.py

Checklist

• Format your code according to the Format code with pre-commit.
• Add unit tests according to the Run and add unit tests.
• Update documentation according to Write documentations.
• Provide accuracy and speed benchmark results according to Test the accuracy and Benchmark the speed.
• Follow the SGLang code style guidance.

Review Process

1. Ping Merge Oncalls to start the PR flow. See the PR Merge Process.
2. Get approvals from CODEOWNERS and other reviewers.
3. Trigger CI tests with comments or contact authorized users to do so.
   • /tag-run-ci-label, /rerun-failed-ci, /tag-and-rerun-ci
4. After green CI and required approvals, ask Merge Oncalls to merge.

[gemini-code-assist]

Warning

You have reached your daily quota limit. Please wait up to 24 hours and I will start processing your requests again!

[gemini-code-assist]

Warning

You have reached your daily quota limit. Please wait up to 24 hours and I will start processing your requests again!

[gemini-code-assist]

Warning

You have reached your daily quota limit. Please wait up to 24 hours and I will start processing your requests again!

[hnyls2002]

Thanks @Johnsonms! This PR edits both python/sglang/jit_kernel/ and top-level sgl-kernel/, neither of which exists any more - they moved to python/sglang/kernels/jit/ and python/sglang/kernels/aot/ (#32072, #32648, #33400) - and the tracking issue #17865 is closed. copy_to_gpu_no_ce is still imported from the AOT wheel at python/sglang/srt/layers/moe/moe_runner/deep_gemm.py:96, so a fresh port would still be welcome. Closing as obsolete - please reopen if I've missed something.