If you want to be pedantic, the JVM does not support tail (function) _calls_, but jumps, which may be the result of tail call optimization or elimination.
As far as I know, we still cannot efficiently turn mutual recursive calls into a chain of jumps, but I'll be glad to be corrected here.
I think there may be a misconception here: A tail call is NOT a call that doesn't allocate a stack frame. A tail call is a call that is the return value of a function. This is why we refer it the stack frame-eliminating optimization as "tail call elimination". We are taking a jump-to-subroutine (call) instruction and replacing it with a normal jump instruction, thus eliminating a "call" from our program.
The JVM most certainly does support general tail calls through its invoke instruction, which supports calls to arbitrary methods of arbitrary objects (which, of course, includes tail calls).
Some compilers also support the optimization of _some_ recursive tail calls (usually recursive tail calls to final or local methods) into loops using the goto instruction, which supports jumps within the current method (essentially optimizing the tail-recursive method into a non-recursive method containing a loop).
If I read you correctly, you say, the JVM 'supports tail calls', because it can call other methods. I think, the phrase 'supports tail calls' is as meaningful as 'supports function calls' then.
In any case, I don't think this is a question of optimization. I don't see how, for example, an F# program written in CPS can ever run on the JVM.
Scheme implementations will compile any tail calls into jumps, even if they're mutually recursive. This is easy to understand by doing CPS transformation by hand on such calls.
As far as I know, we still cannot efficiently turn mutual recursive calls into a chain of jumps, but I'll be glad to be corrected here.