GH-115419: Move setting the instruction pointer to error exit stubs

[markshannon]

Currently, in tier 2, we emit a _SET_IP before any instruction that may error so that the instruction pointer is correctly set.
Since errors are relatively rare, this is wasteful.

With this PR, the error handling uop _ERROR_POP_N sets the instruction pointer, so we don't need to emit the _SET_IP inline.

We still need to emit _SET_IP before instructions that escape.

• Issue: Tier 2 optimizations for 3.13 #115419

[markshannon]

Stats show a ~3% reduction in uops executed overall.

[markshannon]

4 pre-existing failures and a timeout.

[gvanrossum]

The benchmark numbers don't mean anything, do they? IIUC only the stats matter here, and they look good. These benchmark numbers (3% slower) are for the Tier 2 interpreter. If you care to see if it made anything actually faster, run with JIT.

[markshannon]

The benchmark numbers don't mean anything, do they?

No. We are only interested in tier 1 performance for 3.13. For 3.14 we are assuming that the JIT will be working well by then, so we are optimizing for tier uop stats.