One may be familiar with "stack overflow" which is when a series of function calls (often recursive) go so deep before actually returning that it exhausts the maximum stack space (essentially an array of scopes containing variables/state for each function). A proper tail call will realize that it can reuse the same existing stack frame instead of adding a new one, which essentially makes exhausting the stack impossible and removes the need for the CPU to do all of that bookkeeping.
In performance sensitive workflows it can make a large difference.
It sounds like you already know this, but many people, including myself at one time, think of tail call optimization as a trick for not blowing up the stack when writing recursive functions. However, it's much more general. Tail call optimization doesn't have to be recursive, it can be applied any time the final statement or expression of a function is another function call. It's something like a special form of inlining. And as you say, it's actually possible to apply the optimization in limited cases where the tail called function returns a value![2]
That general optimization is very useful for implementing threaded[1] or continuation passing style VMs since it compiles function calls and all their associated baggage down to a jump and maybe some assignments.
[1] Forth style, not multithreading.
I might leave my house, go to an ATM, then go to the grocery store to do my shopping. TCO means that I don't have to stop by the ATM again on the way home.
So my mental model actually doesn't have anything to do with recursion.
If you think about all the standard statements in a language - if, while, async, etc - all of these kinds of constructs can be built out of normal functions once you have tail calls. This makes the language more extensible and reduces the need to introduce new kinds of built-in language constructs and allows more experimentation.
They're really useful for input processing/parsing, recursion, implementing algorithms, etc.
Another way to think about it is that functions can become little 'named blocks' with defined inputs and outputs, and you can combine them together without worrying so much about the runtime costs of functions that take up stack space. So if you have a function with lots of if conditions, while loops, etc, you can convert it into these 'named blocks' instead, which makes maintenance easier.
There are lots of examples. In a way, the fact that none of the main imperative languages support this simple runtime feature is what's stopping people from realising how great tail calls are. There's a bit of a Catch-22 happening.