Make the Pallas launcher dependency-free (torch + jax only)

[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

---

Make the Pallas launcher dependency-free (torch + jax only)

helion/runtime/pallas/launcher.py now imports only torch + jax (+ stdlib) --
no other helion module -- so the precompiler can bulk-export it into a
standalone Pallas kernel with zero Helion runtime dependency. Two helion deps
are severed:

• _PALLAS_UNSUPPORTED_DTYPES (the {int64,uint64,float64} frozenset TPU can't
  lower) moves from helion/_compiler/backend.py -- which only defined it for
  the launcher's import; nothing else used it -- into the launcher.
• is_pallas_interpret (from helion.runtime.settings) is replaced by a
  dep-free _pallas_interpret_enabled() env-var read at the launcher's own
  runtime call sites (the two default_pallas_launcher install-path fallbacks),
  where no CompileEnvironment is active so is_pallas_interpret already
  reduces to that env var (Settings.pallas_interpret itself defaults from it).

_get_vmem_limit_bytes is called both at runtime (launcher) and at compile time
(_compiler/pallas/backend.py, where a CompileEnvironment IS active), so
rather than have it read the env var it now takes interpret as a parameter:
the compiler passes env.settings.pallas_interpret (honoring decorator-set
interpret) and the launcher passes _pallas_interpret_enabled(). This keeps the
launcher dependency-free with no behavior change at either site -- mirroring how
the launcher install path already threads _pallas_interpret as a kwarg.

Verified on the TPU pod: non-interpret pallas add/topk/jax_fn and the
compact-worklist window/budget/guard/matmul tests pass; the launcher imports only
torch+jax. ruff/pyrefly clean (83 baseline); Triton add still correct on
H100. (Interpret-mode add tests fail identically on the base commit -- a
pre-existing pod JAX-version issue, not from this change.)