Ignore non buffer chunk files when resumed

[tagomoris]

This fixes #1098.

buf_file in v0.14 loads all files which matches with glob for buffer file path as queued. That behavior is to read files as much as possible when upgrading Fluentd.
But there is a bug to read files as queued, which is written by out_file in the same directory with buffer chunk files.

This change is to fix not to load files which doesn't have a part like (b|q)[0-9a-f]+, which shows that file is a buffer chunk.

note about #1098

#1098 is a result of combination of this bug and append true, which does read data from buffer chunk file (this file) and also does write these data to output file (this file) - it is infinite loop of read & write.

The minimum steps to reproduce it is:

1. execute command: echo "foobar" > temperature.\ruby -e 'puts Time.now.strftime("%Y%m%d")'.log
2. execute fluentd with configuration below:
3. use ctrl-z to stop Fluentd process without shutdown sequence (shutdown calls flushing)

# no <source> needed
<match demo.temperature>
  @type file
  path temperature
  flush_interval 1s
  append true
</match>

[repeatedly]

Test failed by introducing unknown

[tagomoris]

@repeatedly Oops. I fixed that test (and added another data for pattern newly tested).