Fix scalar-state hashing for PyTorch 1.12 in state hashing utilities

[InsightofSPb]

Motivation

• PyTorch 1.12 handles scalar tensors differently when viewed as raw bytes, so existing cpu.view(torch.uint8).numpy().tobytes() usage could mis-hash or fail for scalar parameters.
• The intent is to preserve deterministic, provenance-friendly raw-byte hashing for both scalar and multidimensional tensors without altering semantics for non-scalar tensors.

Description

• Replace cpu.view(torch.uint8).numpy().tobytes() with cpu.reshape(-1).view(torch.uint8).numpy().tobytes() at the three raw-byte hashing sites to ensure scalars are flattened before byte-viewing in ovs_heritage/stock_features.py and tools/run_official_lposs_parity.py.
• Update module_state_fingerprint to flatten the tensor bytes with reshape(-1).view(...) prior to hashing in ovs_heritage/stock_features.py.
• Add a focused CPU regression test test_scalar_state_hashing_is_deterministic_and_sensitive to ovs_heritage/tests/test_stock_lposs.py that registers a scalar parameter and a multidimensional buffer, verifies deterministic fingerprints, and verifies hashes change when the scalar changes.

Testing

• python -m compileall -q ovs_heritage segmentation/evaluation/lposs_eval.py tools/check_stock_lposs_gpu.py tools/run_official_lposs_parity.py executed and succeeded.
• pytest -q ovs_heritage/tests could not complete collection due to missing runtime dependencies in the environment (errors loading numpy, torch, and yaml).
• ruff check on the modified modules completed successfully and git working-tree checks show the changes staged and committed locally.

---

Codex Task