[cpyrt] Take template arg classes from the type when ranking overloads

[guitargeek]

To rank overloads, CPPMethod::GetPriority() looks up the class of each argument by its name. GetScope() cannot resolve template names like "std::vector" directly, so each such lookup declared a new trampoline in the interpreter, even for a name it had resolved before.

This happens once for every new method with such arguments, e.g. for every specialization of a function template taking containers. The method already knows the argument type, so take the class from the type instead, and force its definition like GetScope() does. Names of std::initializer_list still go through GetScope(), which deliberately resolves them as std::vector: the ranking relies on that (see PR #133 for the initializer_list case).

Add an interop wrapper for Cpp::GetUnderlyingType().

For example:

import time
import cppjit

cppjit.cppdef(r"""
#include <vector>
#include <utility>

template <typename T>
struct Box {
    int probe(const std::vector<T>& v) { return 1; }
    int probe(const std::pair<T, T>& p) { return 2; }
    int probe(int x) { return 3; }
};
""")

types = ["int", "long", "short", "char", "float", "double",
         "unsigned int", "unsigned long"]
extras = [f"std::pair<{a}, {b}>" for a in types for b in ("int", "double")][:12]

vec = cppjit.gbl.std.vector
start = time.perf_counter()
for t in types + extras:
    box = cppjit.gbl.Box[t]()
    box.probe(vec[t]())  # first call per specialization: overload ranking
print(f"first ranked call for {len(types) + len(extras)} specializations: {1000 * (time.perf_counter() - start):.0f} ms")

This takes about 1230 ms instead of 1460 ms (median of 3 trials, Linux x86_64, Python 3.14, CppInterOp/LLVM 22.1.8).

[aaronj0]

We'll have to look at this in the context of the new overload resolution: https://github.com/compiler-research/cppjit/tree/overload-resolution cc @Vipul-Cariappa