Pin the known compression bugs as expected-failure examples

Previously each known bug replaced the affected test for its backend
with an expectedFailure version, so the test as a whole was expected to
fail and nothing checked that the failure was the known one. With
example.xfail() in the stubs (gh-160356), the known inputs are now
expected-failure examples on the affected test: the generated input
leaves the failing region out, as the comments say, and each example
must fail, so the test fails again once the bug is fixed.

For the needs_input bug (gh-159999) the generated calls for bz2 and
zlib are unlimited, which is the case they get right; the examples
depend on where the compressor puts block boundaries, so each has been
checked to fail with both zlib 1.3.2 and zlib-ng 1.3.1, since the
expected failures are strict. The FDICT examples (gh-160001) go on the
zlib wrapper's test_invalid_input, which now clears the flag in generated
garbage rather than skipping the input, and they are only registered in
release builds, where the missing exception is a SystemError rather
than a fatal assertion. The max_length=0 tests for lzma and zstd stop
short of the failing call (gh-160000, gh-160121).

Hypothesis also found that the broad zstd tests could reach gh-160121
with sixteen zero-length calls in a row, so zstd now shares the lzma
carve-out of max_length=0, in the companions and the state machine.

The subTest calls inside the invalid-input and zdict checks are gone:
Hypothesis disables subTest under @given anyway, and inside the stub
replay they swallowed the exception an expected-failure example has to
raise.

Author: pganssle