[BugFix][Relax] Select target-specific pipeline in tvm.compile when GPU target is provided

[swjng]

Problem

relax.build() (exposed as tvm.compile) with relax_pipeline="default" always
resolved to default_build_pipeline, regardless of the target.
default_build_pipeline does not include DLight scheduling — it is a
target-agnostic lowering pipeline. On CUDA, this left TIR functions generated
from ops like Clip/ReLU6 without thread bindings, causing VerifyMemory to fail:

Memory verification failed: Variable `X` is directly accessed by host memory
(it is not contained in a thread environment or in the function arguments).
Did you forget to bind?

Fix

When relax_pipeline="default" and the target is a GPU target
("gpu" in target.keys), use relax.get_default_pipeline(target) which
includes target-aware DLight scheduling. Fall back to default_build_pipeline
if no target-specific pipeline is registered.

CPU targets (llvm, c) continue to use default_build_pipeline unchanged.
The CPU-specific pipeline adds FuseOps/FuseTIR/FoldConstant on top,
which can DCE call_pure_packed calls whose results are unused — correct
per pure semantics, but a separate concern from this fix.

[gemini-code-assist]

Code Review

This pull request updates the Relax VM build process to prefer target-specific default pipelines when the "default" pipeline string is provided and a target is available. A review comment correctly identified that the implementation incorrectly attempts to access the pipeline submodule directly, which would likely trigger an AttributeError and cause an unintended fallback to the generic pipeline; a suggestion was provided to use the properly exported relax.get_default_pipeline function instead.

[swjng]

While investigating the CPU test failures, I noticed that test_vm_compile_simple uses R.call_pure_packed for test.vm.identity, which modifies y's buffer in-place and relies on that side effect:

z = R.call_pure_packed(
    "test.vm.identity", x, y, sinfo_args=(R.Tensor(ndim=2, dtype="float32"))
)
return y  # z unused — relies on y being modified as a side effect

call_pure_packed declares no side effects, so passes like FoldConstant are correct to DCE the call when z is unused. The test should use R.call_packed instead. This didn't surface before because default_build_pipeline doesn't include DCE-capable passes.

I'll send a follow-up PR to fix the test after this merges.

[tlopex]

LGTM