In the bundled zlib-ng, level 1 uses deflate_quick(), which codes all
data with the fixed Huffman codes and never falls back to stored blocks.
Bytes 144 to 255 take 9 bits in these codes, so data made of them with
no repeats grows by 12.5%. Each compressor had an output buffer only
about 10% larger than the block size, so compress_and_crc() raised
OverflowError. At level 1, make the buffer the block size plus an
eighth plus 500 bytes, because no input byte takes more than 9 bits.

When raw.write() raised an error in the thread that writes the
compressed blocks, that thread exited without calling task_done(). As a
result, flush() waited forever, and so did closing the file object that
gzip_ng_threaded.open() returns. write() blocked forever when an input
queue was full.

The thread now stores the error and drops later blocks until stop()
ends it, so write(), flush(), and close() can raise the error in the
calling thread. After an error, close() closes raw only if closefd is
true, so it no longer closes a file object passed to
gzip_ng_threaded.open().

close() now calls flush() before it writes the trailer, as python-isal's
close() has done since pycompression/python-isal@5ef868b, so the trailer
always comes after the queued blocks.

This ports pycompression/python-isal#275.

Update to latest actions

Review feedback: one buffer size for all levels is simpler, and the
extra 2.5% of headroom at the other levels doesn't matter.

Fix `gzip_ng_threaded` writers that raise `OverflowError` at compression level 1

Raise write errors of `gzip_ng_threaded` writers instead of waiting forever