Tail calling (main)

[garrettgu10]

Happy to hop on a call to explain the changes I made, but here's the gist of it:

1. tail_call_funcs.h contains the meat of the implementation, where all the functions are generated using macros. The DISPATCH_GOTO macro was replaced with one that performs tail calls.
2. For instrumented instructions, there was originally a goto that would jump to the portion of the base instruction after the bytecode instruction pointer was incremented. This was replaced with a GO_TO_INSTRUTION macro that subtracts from the bytecode pointer, then tail calls to the original instruction.
3. For error handling, a new CEVAL_GOTO macro was added that first stores the current state to a special struct, then returns an enum (tail call functions return a uintptr_t which is either the returned PyObject* or this enum). On the top level of the _PyEval_EvalFrameDefault, this enum is then matched against the corresponding branch targets within _PyEval_EvalFrameDefault, after restoring the local variables from the special struct (see bottom of tail_call_funcs.h).

I have it building and working on native ARM and WASI, but initial performance tests have been a bit disappointing. On native targets, an easy thing to try would be to build with the latest LLVM 19 and use the preserve_none calling convention to optimize the register allocation and not have to restore register values before tail calling. I care more about Wasm performance, so I will be looking through that first, specifically getting numbers from the V8 side, which is more difficult because of lack of WASI support.