NVFP4: make rich+RSF adaptive scaling default

Enable top1 rich scale search for NVFP4 adaptive quantization with slots
6,5,4,3,2,1.5,1 and ±1 UE4M3 code search. Add RSF-lite per-tensor
scale multiplier fitting by default, plus summary stats for rich candidates,
slot distribution, and RSF multipliers.

Author: mengqin