[Mosaic TPU] Support reshapes of form `(*src_major, src_minor) <-> (*dst_major, dst_minor)`, where `max(src_minor, dst_minor) <= 128` and one divides the other (row merging and splitting).
This change is targeting at small lane dims, since when src_minor/dst_minor are multiples of 128, it is already supported.
PiperOrigin-RevId: 966129068