ATM's cell size is 53 bytes of which 5 bytes are per-cell header.
Higher layer protocols gang together multiple cells through the use of segmentation-and-reassembly. An IP packet would be wrapped in, for example, an AAL5 packet (which itself has a header). That in turn would be segmented and transmitted as cells. The receiver would try to recover from cells arriving out of order, or interleaved with other traffic, or various error conditions, and would deliver fully reassembled AAL5 packets to the higher layer.
One problem is that early ATM switches had small buffers and poor buffering strategies leading to cell loss. The loss of one cell out of an AAL5 frame would require recovery by still higher layer protocols, such as TCP.
Instead of ATM switches (unsuccessfully deployed at the PACBell and Ameritech NAPs in the early 1990s), large ISPs used FDDI switches. FDDI has a token. Switched FDDI exploits the token usefully to impose flow control (it can withhold tokens from aggressive senders), and that made Internet Exchange Points (and indeed LANs internal to ISPs) much more stable than they were using ATM of the day.
The tokens you are thinking about are only relevant on a LAN, just like CSMA/CD does not propagate off a local ethernet to the whole Internet.
Decnet Phase IV is still* in the Internet; it's the transport layer for IS-IS (the routing protocol, not the terrorists). It has other niche applications too.
"I have no idea why the world coalesced on TCP/IP"
It wasn't vendor-specific and ran on arbitrary layer two infrastructure. There were address resolution and mapping systems for local area networks of all sorts (ethernet, starlan, token ring, fddi, and on and on and on), while competing protocols were not adapted to arbitrary links quite as quickly. Additionally, IP was trivially serialized (with SLIP initially, then with HDLC and PPP), while the serializations for vendor-supported protocols were fairly complicated.
In short, the motto "IP on everything" reflected many people's goals.
There were also several reference implementations and interoperability was taken seriously (a regular conference existed to test interoperability, even).
Eventually even IBM machines (from mainframes to AS/400s) learned to talk TCP/IP.