Correct. The parallel only gets you buffering. The CPU is the limiting factor. I do not know if the sensors support retransmission, and I forgot to check the datasheet. If they emit on specific time values they may not.
According to your math you will get 625 samples off the SPI, which means you will lose samples. Even if each is configured for 200 samples a second. If okay with this, using masks will allow round robin of each ID. Note I have not done the math myself.
There could be an issue of CAN bandwidth, but I am thinking this is not a concern. It was technically unknown until the datasheet was added in. However in all fairness this was almost assured. Technically there could be an issue if the data is critical however this is kind of extreme.
The SPI bandwidth is a concern as this will have decent amount of overhead due to the nature of this controller. There is only so many SPI ports available, but the CPU can only be in one place. DMA would change this however it does not increase the SPI ports or RX buffer size.
These probably support ping pong. So in theory you could read one from SPI and CAN write to the other. However this only works if certain conditions are correct. Which is looking doubtful.
Note this is not my project. However I was agreeing with the idea of reprogramming multiple on a single bus. I think the parallel approach is only really needed if there is really critical timing or bandwidth concerns.