Compiler profiling

[mcourteaux]

This is a Proof-of-Concept PR. I tried to be minimally invasive:

1. IRVisitors and IRMutators can be decorated with Profiled<...> to get the version that profiles the visit functions.
2. ZoneScoped macro (and a few variants) times a scope, using the function name in which it sits.
3. Additionally, I added IRVisitor::profiled_visit(), IRMutator::profiled_mutate(), and IRGraphVisitor::profiled_include(), to be sure you can profile the IRVisitor/Mutator without necessarily profiling the entire IR tree being walked.

To use this:

1. Enable in build with CMake with -DWITH_COMPILER_PROFILING. When this is not passed, the code is compiled as before with zero runtime overhead.
2. Run the Halide compiler for some Pipeline with HL_COMPILER_TRACE_FILE=somefile.trace.json as environment variable. If this variable is not set, no profiling events will be collected, nor written.
3. The file will be written when the program exits.
4. you can open this in https://ui.perfetto.dev

Example trace file: vector_shuffle.trace.json

With sample screenshot:

This work led to #8996

[mcourteaux]

@alexreinking Can you validate this change?