The homeservers act as signalling servers. For STUN/TURN you need to run a separate TURN server (typically coturn), as per https://matrix-org.github.io/synapse/latest/turn-howto.html.

The future SFU will similarly be split from the homeserver, with the initial implementation based on either Signal-Calling-Service, ionsfu or mediasoup (we're evaluating all three). Of course, the point of being standards based is that you'll be able to mix & match SFUs and MCUs from other vendors.

You can find out more about the architecture from my talk at Commcon (https://2021.commcon.xyz/talks/extending-matrix-s-e2ee-calls...) - or Robert's talk at FOSDEM: https://fosdem.org/2022/schedule/event/matrix_metaverse/