To put it simply:
All the blockchains, including TON, group new transactions into a "block" before they are added to the chain-like structure that consists of all the previous blocks (hence the word "blockchain"). And it always takes some noticeable time to create a new block, due to a number of factors.
But a block can contain lots of transactions. And that means the time it took to create a block was not spent on a single transaction, but on all of them at once.
So there is no contradiction. When people say "TON can process millions of transactions per second", it means that all that giant stream of transactions won't overwhelm the network. But all the transactions in that giant stream would have that same several-seconds processing time.