Memory planning asserts `graph_output_allocated not set` when a graph output carries no tensors

[msluszniak]

🐛 Describe the bug

Verifier.verify_graph_input_output in exir/memory_planning.py walks the placeholder and output nodes and records whether the planner allocated each one. A node with no tensor specs is skipped, and the skip happens before graph_output_allocated is assigned:

for nd in graph_module.graph.nodes:
    if nd.op in check_list:
        if not (specs := get_node_tensor_specs(nd)):
            continue

So a graph whose output node carries only non-tensor values never sets the flag, and lowering fails on:

AssertionError: graph_output_allocated not set

even though the graph is fully planned. There is nothing to allocate for that node. The neighbouring len(specs) == 0 branch already treats "nothing to allocate" as success for the all-const case.

This is reachable from a normal export. Lowering a llama-family model to Vulkan with dynamic shapes produces a partition whose only outputs are symints (sym_size / et_vk.select_as_symint), and every such export aborts here.

Versions

Reproduces on main (a6b115b).

Repro:

from executorch.examples.models.llama.export_llama_lib import export_llama
from executorch.extension.llm.export.config.llm_config import LlmConfig

cfg = LlmConfig()
# ... qwen3 0.6B checkpoint/params ...
cfg.model.use_kv_cache = True
cfg.model.use_sdpa_with_kv_cache = True
cfg.model.enable_dynamic_shape = True
cfg.backend.vulkan.enabled = True
export_llama(cfg)

A standalone unit repro, no backend needed:

from torch.fx import Graph, GraphModule
from executorch.exir.memory_planning import Verifier

graph = Graph()
sym = graph.placeholder("sym")
graph.output((sym,))
Verifier(
    GraphModule({}, graph),
    alloc_graph_input=True,
    alloc_graph_output=True,
    alloc_mutable_buffers=True,
).verify_graph_input_output()