Re: pg_waldump: support decoding of WAL inside tarfile

I don't like the v3 patches too much: in particular, they do nothing
for the failure-to-finalize bug I identified yesterday. v3-0003
is on the right track but seems overcomplicated. Here is my own
set of proposed patches, which I think will fix what we are seeing
in the buildfarm:

v4-0001 is Andrew's fix for incorrect decompressor finalization.

v4-0002 fixes read_archive_file() to do the missing finalize step.

v4-0003 fixes init_archive_reader() to not depend on cur_file.
This is closely allied to v3-0003 but simpler. I also added
some commentary to pg_waldump.h about what it's safe to do with
cur_file.

get_archive_wal_entry() violates that advice and is pretty much
utterly broken IMO, because it still believes that it can use cur_file
in an incorrect way. However, the impact of that is that it may fail
to flush some hashtable entries out to temp files (in case a single
read_archive_file() step reads more than one WAL file, which is
entirely possible with compression). That is a performance issue but
it's not causing our buildfarm problems, so I left it untouched here.
But I don't think any of the patches proposed so far fix it properly.
What it should do should look more like the revised version of
init_archive_reader's loop: call read_archive_file(), then scan the
hash table for WAL entries we need to flush to files, then finally
return the desired WAL entry if it's present, else loop around.

regards, tom lane