vm: fix compiling ebpf program using llvm in parallel environment

[Officeyutong]

Closes #91

The original compilation implementation has the following bugs:

• More than one thread may execute InitializeNativeTarget in the same time
• Each call to bpf_jit_context::compile() may lead to the creation of a new LLJIT instance, which obsoletes the previous one, and making functions jitted with the previous one being invalidated (the memory they used would be deallocated, causing SIGSEGV when calling them)
• Multiple thread may call bpf_jit_context::compile() for the same program, which can cause data race and resource waste

Fixes are:

• Use an atomic variable to indicate whether the llvm environment was initialized. Only initialize if not initialized previously
• Use a spinlock to control the compilation process. A spin lock will ensure that only one thread could call bpf_jit_context::compile() of a certain bpf_jit_context (In fact, in out implementation, bpf_jit_context::compile() should only be called once per bpf_jit_context, since bpf_jit_context was connected to a unique ebpf_vm, which indicates a ebpf program). If the program was not compiled before, create a LLJIT instance, and compile it, then store the LLJIT instance into bpf_jit_context. Otherwise, directly lookup symbols from the previous stored LLJIT instance.

Besides, this PR updates test-examples workflow, make it enable to use llvm-jit as ebpf runtime