The reason to use hash-chain accumulators rather than Merkle proofs is that with a Merkle tree you need to check 22 hashes, so there’s a ~22x increase in the amount of proving you need to do per value. Also, the implementation complexity goes up considerably (eg. it risks getting to the point where you can’t hand-roll a STARK, you need to use an HLL, which adds overhead and bug surface area…)
My understanding is that at present with reasonably optimized SNARKs and using MIMC you can prove a branch in ~2 seconds. With STARKs you can get that down to ~0.2, and replacing the Merkle branch with a hash chain ~0.01, and probably even less with improved hash functions (Jarvis?) so you get to the point where it actually is viable for a Plasma server.