Preserve clean EOF after non-blocking gzip member

[carrerasdarren-cell]

Summary

• wait for input before starting a possible next gzip member on a non-blocking
  reader
• treat EOF after a completed member as clean when an earlier call observed
  EAGAIN
• add a one-byte-at-a-time pipe regression for the complete-stream case

Root cause

After Z_STREAM_END, gz_look() immediately switched from LOOK to GZIP
when no compressed input was buffered. On a non-blocking descriptor, an
EAGAIN after the completed trailer therefore committed the state machine to
a next member that might never arrive. If EOF appeared on a later call, the
empty next member was reported as truncated.

The complete uncompressed payload was returned, but gzclose() returned
Z_BUF_ERROR. The regression reproduces this with public APIs and a local
pipe: it feeds one valid member one byte at a time, drains to EAGAIN after
each byte, and closes the writer only after the trailer has been consumed.

The fix asks gz_avail() for input while still in LOOK. An EAGAIN leaves
the reader ready for a delayed concatenated member, while a later EOF with no
bytes is clean.

This is distinct from the open non-blocking write-side reports and patches in
#1256, #1281, #1286, and #1292.

Validation

• focused regression fails on untouched develop with
  gzclose returned -5 after a complete stream
• patched Clang ASan/UBSan CMake suite: 12/12 pass
• traditional static and shared make test pass
• strict C89 CMake suite: 11/11 pass
• focused GCC 15 -Wall -Wextra -Werror build and test pass
• additional local sequence coverage preserves delayed concatenated members,
  transparent reads, trailing garbage handling, and Z_BUF_ERROR for a truly
  truncated trailer