Commpage: Process-Isolated Benchmarking via Shared Memory

[bazhenov]

Motivation

Today tango loads both the baseline (B) and candidate (C) benchmarks into the same process using dlopen/LoadLibrary. This is clever but has drawbacks:

1. Platform hacks -- Linux PIE patching (linux.rs) and Windows IAT patching (windows.rs) exist solely to make executables loadable as shared libraries. These are fragile and there is no guarantee they will continue to work in the future. Especially on Linux where this approach was explicitly banned by dyld.
2. Insufficient isolation -- Operating systems provide ways to collect additional per-process information (e.g. memory allocations), but these are often defined at the process level, not the thread level. Running both benchmarks in the same process prevents effective use of these tools.

Running each benchmark in its own process and reading measurements through shared memory (the "commpage") solves both problems.