Test: Count a TSan report only when the racing access is in the extension - #230

The first-report output added in #221 showed what failed on Linux CI: two PyUnicode_FromString memcpys in libpython, writing into memory that the uninstrumented allocator had handed from one finished thread to another. TSan cannot see that synchronisation. Our tp_wrap_usek only appeared further up as the caller. The old filter counted any report with an extension frame anywhere in it.

Now a report counts only when the access itself was made by the extension: the first frame of either racing access outside the sanitizer runtime must be in ftmod.

Verification:

• On the CI report from the Test: Show the first TSan report on failure; the plain-epoch TSan check is a race mutant #221 run, the new filter does not count it.
• A synthetic report with tp_* accesses in ftmod is counted.
• With real TSan locally (3.15.0rc1t), both TSan tests pass. That includes the race-mutant test_threadsanitizer_sees_a_plain_epoch, so a planted race in the header is still caught.