That term was a long time ago already seen as problematic in the Lisp community. At the time when Ousterhout used it for marketing TCL.
Basically a scripting language interpreter runs source code files, called scripts.
Lisp does that, too. But many Lisp implementations compile the code and often to machine code, since the runtime includes an incremental compiler. The incremental compiler can compile individual expressions to memory.
Thus when you use a Lisp system, running a script just looks like running an interpreter, even though the code gets compiled incrementally.
Lisp uses the function LOAD, which can load a textual source file and which can then read, compile, execute expressions from the file. Typically this is also available via command line arguments.
Like in SBCL one can do:
sbcl --script foo.lisp
and it will read/compile/execute expression by expression from a textual source code file, actually incrementally compiling each expression to machine code.
I chose this based off a random search on google of "scripting language." You can do the same and would get the same result.
It's not a rigorous term, because (you are right) its a property of the implementation not the language itself, despite the name.
But I don't think Julia has relevance to discussion of the speeds of interpreted Lisp.