Move Triton launcher into helion/runtime/triton/launcher.py

[AmesingFlank]

Stacked PRs:

• docs: update precompile docs for pretuned_kernel, jax_fn, helper embedding #3189
• Rename aot_kernel to pretuned_kernel #3188
• Embed in-kernel Pallas runtime helpers into generated code #3187
• Support OutputCodeOptions(jax_fn=True) for Pallas (JAX-array entrypoints) #3186
• Refactor the Pallas launcher to make it easier to operate on Jax Arrays as well as TorchTPU tensors #3190
• docs: document helion.precompile for standalone kernels #3185
• Support OutputCodeOptions(allow_helion_deps=False) for the Pallas backend (TorchTPU) #3184
• Make the Pallas launcher dependency-free (torch + jax only) #3183
• Move Pallas launchers into helion/runtime/pallas/launcher.py #3182
• Add OutputCodeOptions(allow_helion_deps=False) to BoundKernel.to_code for standalone Triton kernels #3181
• Relocate get_num_xcd into the dependency-free Triton launcher #3180
• Make the Triton launcher helion-dependency-free (torch + triton only) #3179
• ->Move Triton launcher into helion/runtime/triton/launcher.py #3178

---

Move Triton launcher into helion/runtime/triton/launcher.py

Relocate the runtime symbols that Helion's generated Triton code depends on
at execution time out of the catch-all helion/runtime/__init__.py into a
dedicated helion/runtime/triton/launcher.py:

• default_launcher — runs a compiled triton.jit kernel
• get_num_sm — persistent-kernel grid size (host statement)
• set_triton_allocator (+ _alloc_fn) — TMA / tensor-descriptor scratch allocator

helion.runtime re-exports all three, so helion.runtime.default_launcher,
helion.runtime.get_num_sm, helion.runtime.set_triton_allocator, and the
generated from helion.runtime import default_launcher are unchanged.

This is a pure move with no behavior change — the module still reaches back into
a few helion helpers (exc, settings, _utils). Isolating these into one
small module is the first step toward making them dependency-free, so the
ahead-of-time precompiler can bulk-export exactly this file into a standalone
kernel.

Because get_num_sm now resolves its globals in the new module,
test/test_runtime.py is updated to patch
helion.runtime.triton.launcher._module_is_pallas_interpret (the moved
function's real lookup site) instead of helion.runtime._module_is_pallas_interpret.

Verified: ruff/pyrefly clean, test_runtime and test_persistent_kernels
pass, and the generated Triton for examples/add.py is byte-identical.