Add guarded long-context benchmark tooling

[francescogiunta]

Summary

• add a Linux process-group memory guard for model-backed benchmark runs
• add an optional decode warmup window while preserving the existing CSV columns and meanings
• add a deterministic fact-recall fixture sized for 65K and 100K context frontiers
• add synthetic tests and usage documentation for the new tooling

Motivation

Long-context measurements on unified-memory systems need an explicit safety envelope and a reproducible workload. Decode throughput also benefits from reporting a post-warmup window without changing existing benchmark metrics.

Compatibility

The warmup defaults to zero. The original nine CSV columns remain in their existing order and keep their current meaning; the warmup and measured-window fields are appended. The canonical story fixture is not modified.

Validation

• make test-memory-guard — 7 tests passed
• make test-bench-metrics — C metrics and CLI tests passed
• deterministic fixture regeneration produced SHA-256 31cb363cd5b5b7eaccffdf9c464ba491d6afea46b54b4d5baf3dde36299d7436
• git diff --check passed

No model-backed performance result is claimed in this draft; those runs remain guarded campaign work.