Benchmarking open-source LLMs on Verilog RTL generation across 50 tasks