PoW chains can theoretically be configured to target any arbitrary block time but the block time is inversely proportional to the orphan rate (rate of unincluded valid blocks due to network latency). The higher the orphan rate, the lower the economic incentives to mine because more work is left unrewarded, the lower the network hashrate, and therefore the lower the security of PoW.
Ethereum solved this problem via the introduction of “uncle blocks”, basically issuing an extra reward to orphaned blocks in addition to included blocks. This is how Ethereum can reach a target block time of 15 seconds instead of Bitcoin’s 10 minutes. This does however force the community of ETH tokenholders to subsidize the cost of security through extra inflation. If we decrease the block time even further, sub-10s as you proposed, we would have to issue rewards for more orphaned blocks and subsidize the security of the network via more inflation, driving ETH value down over time.
To answer your original question about high throughput BFT consensus algorithms (e.g. Tendermint), byzantine fault-tolerant finality is based on the assumption that at least 2/3+1 of preset validator nodes will remain online and honest at all times. Because the number of validator nodes is preset, finality is achieved once a required quorum of validator signatures is present. There is no need for further block confirmations because of the 2/3+1 assumption.
Additionally, since we’re already making the 2/3+1 assumption, non-validator nodes generally do not have to run full nodes or download full blocks to verify the validity of the chain because they rely can query information about the blockchain from the validators based on the assumption that 2/3+1 of them are honest.