Susbstantial optimization to deinstrument/re-instrument

[xirafa]

I've seen an experimental bytecode deinsumentation using "--immediate".

It didn't really work for me on 3.11, then I've looked into it and apparently python 3.11 maintains two versions of the bytecode, (the second version is the "adaptive" version per PEP659, and lies in memory just after the code object in CPython) - both of which need modifying. I added deinstrumentation and performance has skyrocketed, without the need to maintains stats and heuristic for slower deistrumentation.

This approach also enables re-instrumentation for measuring individual test coverage.

The POC of this is unfortunately proprietary but it could be huge win if implemented in slipcover.