Align GPTQ-Pro kernel with Ampere GPUs (auto-selection + native SASS)

[Copilot]

• Raise SUPPORTS_FORMATS priority 0 → 95 so GPTQ-Pro enters auto-selection on Ampere
• Add -gencode flags for sm_80/sm_86/sm_87 (native Ampere SASS)
• Add -gencode arch=compute_87,code=compute_87 PTX fallback for Ada/Hopper forward compatibility

[groxaxo]

@copilot review the changes please.

[Copilot]

@copilot review the changes please.

Here's the review of all changes in this PR:

Commit 1 — QUANTIZE_BLACK_LIST (a81fc20)

• Correct. QUANTIZE_BLACK_LIST was an empty dict {}, so no method was ever blocked. The change correctly lists non-GPTQ methods (GGUF, FP8, BitsAndBytes, QQQ, AWQ, EXL3, PARO) and improves the error message to tell the user what method to use instead.

Commit 2 — Ampere alignment (39e5aae)

• SUPPORTS_FORMATS priority 0 → 95 — Correct. Priority 0 permanently excluded GPTQ-Pro from BACKEND.AUTO selection (the selector skips anything with priority <= 0). At 95 it ranks above Marlin (90) and ExllamaV2 (80). The existing validate_device() guard (major >= 8) ensures transparent fallthrough to Marlin on pre-Ampere hardware.
• -gencode flags — Correct, with one issue I found and fixed: _validate_gptq_pro_device_support() checks major >= 8, which also passes on Ada Lovelace (sm_89 / RTX 4090) and Hopper (sm_90 / H100). With only SASS cubins for sm_80/86/87 and no PTX fallback, loading the kernel on those devices would fail with "no kernel image available". Fixed in 0457356 by adding -gencode arch=compute_87,code=compute_87, which embeds sm_87 PTX that the CUDA driver can forward-JIT for any post-Ampere device.