Ttimms/Spark-X2.5-4B-NVFP4

Swap RTN -> GPTQ build: HumanEval -4.3pp / MBPP -2.0pp vs bf16 (RTN was -8.5 / -15.6); finalize card with 3-way table

docs: add HumanEval eval (RTN -8.5pp vs bf16); GPTQ + MBPP in progress

NVFP4 W4A16 (RTN) of Spark-X2.5-4B; coherence-verified, HumanEval/MBPP + GPTQ to follow

initial commit