[executorch][native] Add Graph index arena to the in-memory IR

[SS-JIA]

Stack from ghstack (oldest at bottom):

• [executorch][native] Add Method + ValueRole to the in-memory IR #22281
• -> [executorch][native] Add Graph index arena to the in-memory IR #22280
• [executorch][native] Add Node to the in-memory IR #22279
• [executorch][native] Add Argument tagged union to the in-memory IR #22278
• [executorch][native] Add Value + index-arena ids to the in-memory IR #22277
• [executorch][native] Add Scalar value type to the in-memory IR #22276
• [executorch][native] Add TensorMeta + ScalarType in-memory IR types #22275
• [executorch][native] Add native runtime Program reader + DOT visualizer #22274

Adds ptn::Graph, the index arena that owns the value layer landed so far. It
is a pure function body (mirrors the schema Graph): it holds the Nodes and
Values that the NodeId / ValueId handles point into, the ordered graph
input / output value lists, and -- recursively -- the subgraph bodies for
higher-order ops. Stateful method-level bindings (constants, output specs,
mutable buffers) are deferred to a later Method type.

Design decisions baked in:

• Per-graph subgraph arena. Graph owns std::vector<Graph> subgraphs and
  GraphId indexes the enclosing graph's subgraphs, matching the schema's
  recursion and the per-Graph SSA namespace so a subgraph stays self-contained
  with its parent. (A std::vector<Graph> member of Graph is legal C++17 --
  the standard containers permit an incomplete value type at the point of the
  member declaration.)

• Storage identity vs execution order are decoupled. nodes is an
  append-only arena so a NodeId never shifts (the index-arena invariant),
  while schedule (std::vector<NodeId>) carries the topological / execution
  order a runtime walks. At load the arena order equals the wire's topological
  order and schedule is the identity [0, n) (initialize_schedule()); across
  mutation the arena order is no longer topological, so schedule is
  authoritative -- reorder / insert there (moving int32s, invalidating no id)
  rather than moving storage. Deletion via tombstone + a compacting pass is
  deferred until a mutating pass needs it.

• Pure arena. inputs / outputs are ValueId lists (schema SSA names
  resolved to ids at deserialize); no tensor_values side table in memory
  (each Value already carries its TensorMeta); the name to id map stays
  deserializer-local.

rebuild_def_use() recomputes every Value's producer / consumers from the
nodes (order-independent -- it walks the arena, not schedule). Placeholder and
Output nodes are real entries in nodes, so def-use is uniform: a graph input
value's producer is its placeholder node, and graph inputs are identified by
membership in inputs, not by producer == kInvalid. This corrects the now
stale Value.h producer comment (also in this diff). Bounds-checked
node() / value() / subgraph() accessors throw on an invalid id, and the
IR printer gains to_string(const Graph&), a multi-line dump in schedule
order.

consumer_ids is a set of consuming nodes rather than a bag of uses, so
add(x, x) lists its consumer once and size() counts consumers. Since nodes
are walked in arena order a repeated operand appends consecutively, so a tail
check keeps that exact without a lookup structure.

rebuild_def_use() distinguishes the two things an unusable id can mean.
kInvalid is an absent operand and is skipped; an id that is set but does not
address the value arena can only be a corrupt graph, and now throws rather than
being skipped, which would have left def-use half-wired with no signal. That
matches the accessors, which already throw on a bad id.

Pure std only (no ExecuTorch, no flatbuffers), consistent with the rest of the
standalone ptn runtime.

Differential Revision: D114396767

[pytorch-bot]

🔗 Helpful Links

🧪 See artifacts and rendered test results at hud.pytorch.org/pr/pytorch/executorch/22280

• 📄 Preview Python docs built from this PR

Note: Links to docs will display an error until the docs builds have been completed.

✅ You can merge normally! (1 Unrelated Failure)

As of commit 50bc327 with merge base a5f15b5 ():

FLAKY - The following job failed but was likely due to flakiness present on trunk:

• pull / test-qnn-testsuite-linux / test-backend-linux (qnn, models) / linux-job (gh) (detected as infra flaky with no log or failing log classifier)

This comment was automatically generated by Dr. CI and updates every 15 minutes.

[github-actions]

This PR needs a release notes: label

If your change should be included in the release notes (i.e. would users of this library care about this change?), please use a label starting with release notes:. This helps us keep track and include your important work in the next release notes.

To add a label, you can comment to pytorchbot, for example
@pytorchbot label "release notes: none"

For more information, see
https://github.com/pytorch/pytorch/wiki/PyTorch-AutoLabel-Bot#why-categorize-for-release-notes-and-how-does-it-work.

[digantdesai]

Review automatically exported from Phabricator review in Meta.