bpo-43760: Speed up check for tracing in interpreter dispatch

[markshannon]

The code to check if tracing is active occurs once per opcode in the interpreter and is executed in the order of 100 million times per second. It needs to be as efficient as possible. It cannot be kept in a register as it needs to be modified outside of the interpreter, so on the C stack is the fastest possible location for it.

This PR keeps use_tracing in the C frame of the current interpreter making the check for tracing as fast as possible.

Doing so complicates access to use_tracing from other code and requires us to be careful about maintaining stack discipline.
For anything less performance critical than use_tracing this probably wouldn't be worth it, but as use_tracing is so performance critical it is worth it.

https://bugs.python.org/issue43760

[bedevere-bot]

🤖 New build scheduled with the buildbot fleet by @markshannon for commit 5ba8650 🤖

If you want to schedule another build, you need to add the ":hammer: test-with-buildbots" label again.

[bedevere-bot]

@markshannon: Please replace # with GH- in the commit message next time. Thanks!