perf: cache the lookup of should_trace_cache

[nedbat]

Doesn't seem to improve performance by more than about .5%

pyver	proj	without cache	with cache	compare
python3.11	bug1339.py	0.767 s	0.819 s	106.77%
python3.11	bm_sudoku.py	50.346 s	50.142 s	99.60%
python3.11	bm_spectral_norm.py	60.180 s	59.727 s	99.25%

[nedbat]

This is about what I expected, since the dict lookup is only once per function call, and dict lookups are fast.