Quantized GEMM

quantized_gemm exercises packed low-precision values, dequantization-like math, and projection-style arithmetic relevant to quantized inference paths.

pantheon --test quantized_gemm --duration 60 --gpu 0 --mem 50 --verify --profile

Area	What it tests
Precision path	Packed low-precision unpacking and scaling proxy.
Compute	Dense projection-style arithmetic.
Memory	Repeated reads through a synthetic weight/context working set.
Result	Synthetic quantized operations per second.

Use it for comparative hardware diagnostics. It does not claim a specific INT4, INT8, FP8, or FP4 model-kernel implementation or accuracy result.