Is that true? I think of C# and Java as compiled languages and they run on the CLR/JVM which are both virtual machines.

Compiling to machine code rather than VM bytecode is a difference in degree rather than kind, and doing it well rather than naively is a deep topic.

Simple machine code generation with basic register allocation and evaluation which eagerly spills to the stack is not super hard. Producing fast code instead is a bottomless well.

There's literally no difference in the theory or technique of compiling for a virtual machine or a real one. Real machines are just messier and have a bunch of edge cases and weirdness that require you to read those multi-kilopage manuals to figure out. Having one or more intermediate steps is useful for portability and optimization though, so it's probably a good idea to target a virtual machine unless you only intend to ever target one specific architecture that you know inside and out.

Plot twist - the developer then runs the code in qemu for rapid iteration and ease of debugging... ;)

Generating code is not all that hard. Generating fast code is a challenge. Decoding instructions on some architectures is hard. If you have a nice general IR naively generating native code can be just a table lookup with a register tracking/allocation routine.