SpeContext: Enabling Efficient Long-context Reasoning with Speculative Context Sparsity in LLMs