Embed in-kernel Pallas runtime helpers into generated code

[AmesingFlank]

Stacked PRs:

• Support OutputCodeOptions(jax_fn=True) for Pallas (JAX-array entrypoints) #3186
• Support OutputCodeOptions(allow_helion_deps=False) for the Pallas backend (TorchTPU) #3184
• Add OutputCodeOptions(allow_helion_deps=False) to BoundKernel.to_code for standalone Triton kernels #3181
• ->Embed in-kernel Pallas runtime helpers into generated code #3187
• Refactor the Pallas launcher to make it easier to operate on Jax Arrays as well as TorchTPU tensors #3190
• Make the Pallas launcher dependency-free (torch + jax only) #3183

---

Embed in-kernel Pallas runtime helpers into generated code

The generated Pallas kernel calls two helion-defined, pure-jax helpers --
divide_filter_topk (aten.topk lowering, via _helion_divide_filter_topk) and
flatten_worklist (compact-worklist builder) -- which were pulled in with
from helion... import ..., leaving a helion dependency that blocks
precompiling to a helion-free standalone.

Add a Backend.embedded_helper_source(body) hook (default "") that
BoundKernel.to_code injects between the imports and the kernel body. The
Pallas backend overrides it to inline the source of whichever helper body
references -- module docstring and from __future__ stripped by _embed_source
(located via ast, so a docstring containing a triple-quote can't corrupt the
output) -- plus a _helion_divide_filter_topk = divide_filter_topk alias. The
topk library_imports entry and the inline flatten_worklist import in
render_build_worklist are dropped. Both helper modules import only
jax/stdlib (and now carry a note to stay free of any helion import, which
would trip the precompiler's substring guard), so the embedded module stays
dependency-free. This runs for regular kernels too, keeping a single code path.

Tests: the two builder tests that exec the rendered _build_worklist now
supply flatten_worklist (no longer imported inline);
test_matching_kernel_generates_compact and test_topk_divide_and_filter_lowering
assert the embedded def + alias are present, the old helion import is gone, and
the module still ast.parses (guarding against a corrupt embed). Also repairs
the window-guard test's stale call to the 2-arg _get_vmem_limit_bytes.

Verified on TPU: topk (3 passed), compact guard/render/builder (10 passed); full
test_pallas_compact_worklist.py shows only the 6 pre-existing scratch_types
JAX-skew failures (identical to base), 66 passed. ruff/pyrefly clean (83
baseline); H100 Triton unaffected (embedded_helper_source is a no-op there).

[AmesingFlank]

Folded into #3184. In-kernel Pallas helper embedding (divide_filter_topk / flatten_worklist) is now a dependency-free-only step inside to_code(allow_helion_deps=False) (via embedded_helper_source, called only by the standalone builders) rather than an always-on pass over every to_code. Regular output imports the helpers from helion; only standalones embed them. The helper-module notes and the _get_vmem_limit_bytes test fix moved into #3184, so this standalone commit is no longer needed.