Delay errors in semanal for proper unreachability information

[A5rocks]

• Fixes unexpected attr-defined error in imports in dead code #20408
• Fixes Platform check is not remembered after return #14940

This is a draft PR because:

• it doesn't work with dmypy
• it's a bit slower:

=== Results ===

master                    2.697s (0.0%) | stdev 0.012s
f18e550f23d30c454a0f6fc2c8b41eb9e03b4f29 2.732s (+1.3%) | stdev 0.018s

[A5rocks]

Idea for a speedup (so I don't forget): only do this for things that can be skipped unreachable (I.e. stmts in a block, toplevel stmts).

Also note that I haven't cleaned this up, which would involve removing eg the assert False special case and adding a test case for the issue this fixes.

[A5rocks]

I made the changes I thought would help but this is still slow:

=== Results ===

master                    2.722s (0.0%) | stdev 0.030s
806266c53875d8891cfeff412e75dafbd33faa30 2.757s (+1.3%) | stdev 0.032s
Total time taken by the whole benchmarking program (including any setup): 4 minutes, 18 seconds

I suppose this is just going to be slow no matter what...

Second run with better stdev (still slower):

=== Results ===

master                    2.700s (0.0%) | stdev 0.015s
806266c53875d8891cfeff412e75dafbd33faa30 2.732s (+1.2%) | stdev 0.020s
Total time taken by the whole benchmarking program (including any setup): 4 minutes, 29 seconds

[A5rocks]

Now it's slightly faster:

=== Results ===

master                    2.707s (0.0%) | stdev 0.012s
67abbb30                  2.722s (+0.6%) | stdev 0.028s
Total time taken by the whole benchmarking program (including any setup): 4 minutes, 22 seconds

I'm going to assume that's an acceptable slowdown. If not, I'm out of ideas on how to make it faster. I still have to clean this PR up and add test cases, but I will mark it as ready for review because the logic is finished! (I haven't put aside the time to look at dmypy yet...)

---

mypy_primer looks fine except the unreachability thing, I also need to check that out.

[github-actions]

Diff from mypy_primer, showing the effect of this PR on open source code:

strawberry (https://github.com/strawberry-graphql/strawberry)
+ strawberry/relay/types.py:846: error: Unused "type: ignore" comment  [unused-ignore]

comtypes (https://github.com/enthought/comtypes)
- comtypes/test/test_server.py:326: error: Module "comtypes.gen" has no attribute "TestLib"  [attr-defined]

egglog-python (https://github.com/egraphs-good/egglog-python)
+ python/tests/test_unstable_fn.py:177: error: Unused "type: ignore" comment  [unused-ignore]

Expression (https://github.com/cognitedata/Expression)
- README.py:498: error: Name "z" already defined on line 496  [no-redef]
- README.py:572: error: Name "z" already defined on line 570  [no-redef]
- README.py:628: error: Name "xs" already defined on line 211  [no-redef]
- README.py:632: error: Name "mapper" already defined on line 207  [no-redef]
- README.py:633: error: Name "predicate" already defined on line 208  [no-redef]
- README.py:634: error: Name "folder" already defined on line 209  [no-redef]

AutoSplit (https://github.com/Toufool/AutoSplit)
- src/utils.py:147:12: error: Name "win32ui" is not defined  [name-defined]
- src/utils.py:155:29: error: Name "ctypes" is not defined  [name-defined]
- src/utils.py:156:5: error: Name "ctypes" is not defined  [name-defined]
- src/utils.py:159:9: error: Name "ctypes" is not defined  [name-defined]
- src/utils.py:160:9: error: Name "ctypes" is not defined  [name-defined]
- src/utils.py:163:19: error: Name "win32gui" is not defined  [name-defined]
- src/utils.py:176:20: error: Name "FilterGraph" is not defined  [name-defined]
- src/utils.py:182:12: error: Name "COMError" is not defined  [name-defined]