Handle missing AgentFlan query for string suffixes

[taking-lying-flat]

Summary

• allow calculate_loss_scale to receive a missing query
• skip AgentFlan query-map matching when the current context has no query

Root cause

String template suffixes are processed as assistant contexts, but only response contexts receive the current query. AgentFlan therefore calls calculate_loss_scale with query=None for suffixes. The query-map lookup attempted key in query without guarding the optional value, raising a TypeError during template encoding.

Impact

AgentFlan loss scaling can process string suffixes such as Qwen ChatML's <|im_end|>\n while preserving query-specific loss scaling for response contexts.