Assessing Evasion and Behavioral Trade-offs of Obfuscation-Conditioned LLMs in PowerShell Attack Synthesis