Summary

Fix Triton dequant_kernel for GPTQ INT3 (bits=3):

• Remove invalid tensor-valued Python if control flow (Triton compile error).
• Align 3-bit dequantization with GPTQ CPU packing format (10-1-10-1-10, 32 values → 3 int32 words).

This makes Triton INT3 inference compile correctly and produces numerically consistent results with the CPU reference.

Related Issue

Fixes / relates to: #2251

Notes

• This is a correctness-first fix: it restores functional INT3 Triton inference and matches the CPU reference PackableQuantLinear / pack_block logic for 3-bit GPTQ. Performance has not been fully optimized or benchmarked yet.
• Tested on INT3 GPTQ models (e.g. Qwen3-4B) with wikitext; perplexity returns to a reasonable range.