fix(graph): preserve recurrent graph inputs and allocation lifetimes

[big-hip]

Graph capture can reuse an allocator block without freezing it, trim reactivated graph storage, or omit reduction/scalar-power/MetaX concatenation operations from replay. These defects break indexed recurrent inference. The three commits repair allocation ownership, reduction recording, and MetaX last-axis concatenation separately using existing graph interfaces.

Branch: fix/graph-replay-lifetimes. Base: InfiniLM-v0.2.9c.

The diff contains eight files (+185/-39), including focused allocator, changed-input reduction, axis-cache and strided-concatenation regressions. It adds no graph compiler or Prefill capture feature.

Validation: current C500 slice tests passed for reduction/scalar-power replay (5) and concatenation (6); the same run separately checked the precision PR (2). Previous matching-source validation covered NVIDIA concatenation, reduction and allocator behavior. The current production source is byte-identical to that validated implementation. Format and whitespace checks passed. The new split branch has not undergone a fresh standalone accelerator build; combined-runtime tests do not replace that check.

• Minimal scoped commits and regression coverage.
• Formatting and whitespace checks.
• Fresh build of the isolated branch and upstream CI.

Published as a draft while independent branch builds and upstream CI are pending.

Saved output, source provenance, device conditions and archived measurements: immutable evidence. Attachments live on a separate fork branch and are not part of this source diff.

CI at publication: fork Ruff passed; the independent fork CPU build/test matrix is in progress. The upstream run reports action_required and needs repository-side approval. No successful upstream build is claimed.