In the free-threaded build, an atomic load is needed to safely read tp_flags, avoiding data races. Replace bit tests on tp->tp_flags with calls to _PyType_HasFeature(). When multiple bits are being returned as a result of the test, use PyType_GetFlags().

I think some of these flag tests can safely be a normal load rather than an atomic load. However, it seems better to consistently use _PyType_HasFeature() unless we are sure there is no data race possible.

Implementation notes:

• this is unfortunately quite a lot of code churn because there are a lot of places we read tp_flags directly. I used this command to find all the places: rg -tc -- '->tp_flags \&' and then used editor macros to make the changes.
• the flag test expression and _PyType_HasFeature() are not always equivalent since the function only returns 0 and 1 while the test expression might care about individual bits. I wrote a replacement _PyType_HasFeature function that checks how many bits are being passed as the feature and then ensured that only one bit is set. I then manually fixed or confirmed cases were this wasn't so. So, I'm fairly sure using _PyType_HasFeature() as this PR does is correct in all cases.
• I haven't done a performance comparison on a weakly ordered CPU (e.g. ARM). It's possible this change could negatively impact performance. I'll schedule a pyperformance comparison and link it here.

Related PRs: gh-120210 gh-127588

• Issue: Data race on tp_flags with subinterpreters and static types #129817