xar: clamp the input count before the 32 bit avail_in fields

The XAR reader gets its byte count from __archive_read_ahead(). With
archive_read_open_memory() that count can exceed 4 GiB, because the
whole buffer arrives as one block.

The gzip and bzip2 branches cast the count into avail_in, which is 32
bits wide. A count of exactly 4 GiB truncates to zero. inflate() then
returns Z_BUF_ERROR, and the read fails with "File decompression failed
(-5)". The bzip2 branch truncates the same way into bzstream.avail_in.

Clamp both counts. This is the same guard that the 7-Zip reader already
applies to the same fields. The clamp changes the local avail_in and
avail_out, so the *used and *outbytes counts that both branches report
from those locals stay correct.

The lzma branch needs no guard, because lzma_stream holds both counts
in a size_t.

To reproduce: build a XAR archive with one gzip encoded entry whose
declared compressed length is 2^32, put a short zlib stream at the head
of the heap, pad the heap out to that length, then call
archive_read_open_memory() and archive_read_data_block(). Before this
change the read fails with "File decompression failed (-5)".