The most practical uses of recursive functions are those that cannot be easily eliminated. For an interpreter, it could mean dozens of functions all calling each other in a pattern determined at runtime.
The issue used to be with LLVM, yes, but that appears to be solved (see Clang's 'musttail' attribute). Instead it's more about language design. Constructors, destructors, and all of the fancy language features Rust has must work with this new feature, with no changes being visible elsewhere.
Tail call optimization isn’t just for recursive functions. Any tail call can be optimized to little more than a jump, which makes implementing threaded code or similar much nicer. Your compiled code will just push the instruction pointer along an array of opcodes in a tight loop with no function call overhead.
The issue used to be with LLVM, yes, but that appears to be solved (see Clang's 'musttail' attribute). Instead it's more about language design. Constructors, destructors, and all of the fancy language features Rust has must work with this new feature, with no changes being visible elsewhere.