Condition-Invariant fMRI Decoding of Speech Intelligibility with Deep State Space Model