fix(ingest): guard topcpu DX12 scores by GPU architecture

[Seungpyo1007]

The topcpu GPU resolver previously only documented a release year >= 2011 guard, leaving enrichment dependent on a fragile year filter. TeraScale and DX12-capable GCN cards shipped in overlapping years, so recent dates could admit extrapolated Time Spy results for unsupported hardware.

Pass each TechAPI record's architecture into the Time Spy and GPU breadth resolvers, and suppress Time Spy, Time Spy Extreme, and Speed Way scores for known pre-DX12 architecture families, including versioned TeraScale/Tesla labels and older ATI/NVIDIA families. Preserve non-DX12 fields, unknown-architecture behavior, exact model matching, and fill-only-nulls; no TechAPI records are modified or backfilled. This guard excludes pre-DX12 hardware and does not implement the stricter, separate Speed Way/Ultimate eligibility rules.

Architecture boundaries were checked against AMD's Windows driver support matrix, NVIDIA's DirectX support documentation, and UL's Time Spy requirements. Fermi (including 2.0) and early Kepler remain eligible for Time Spy: they support the DX12 API, and Time Spy requires feature level 11_0 rather than 12_0.

Validation: full pytest suite passed (578 tests; 77.69% coverage, above the 60% CI threshold), mypy app passed, and ruff check app tests passed. Pytest used a fresh --basetemp directory to avoid an existing Windows temp-directory permission issue. Regression coverage includes unsupported TeraScale/Tesla, capable GCN/Fermi/Kepler, overlapping release years, misleading dates, Tesla-branded Kepler hardware, both enrichment entry points, and preservation of non-DX12 dimensions.

Refs #99