Fix race condition of retry state in flush thread

[repeatedly]

This is simple solution but performance is bit decreased.
The better solution is changing retry mechanizm, e.g. don't share state, but it makes
the code complicated. This is feature task.

This fixes #1555

[repeatedly]

Reproduce step:

• conf

# dummy plugin is also ok. Need to generate lots of events.
<source>
  @type tail
  path /pat/to/log
  tag test.tail
  read_from_head true
  <parse>
    @type none
  </parse>
</source>

<match test.**>
  @type buf_test
  @id buf_output

  <buffer>
    chunk_limit_size 500000
    flush_interval 5s
    flush_thread_count 16
    #retry_max_times 15
    retry_max_interval 2s
  </buffer>
</match>

• out_buf_test.rb

require 'fluent/output'

module Fluent::Plugin
  class BufTestOutput < Output
    Fluent::Plugin.register_output('buf_test', self)

    config_section :buffer do
      config_set_default :chunk_keys, ["tag"]
    end

    def initialize
      super
      @num = 0
      @m = Mutex.new
    end

    def configure(conf)
      super
    end

    def multi_workers_ready?
      true
    end

    def start
      super
    end

    def close
      log.error @num
      super
    end

    def write(chunk)
      if rand(10) > 7
        raise
      end

      @m.synchronize {
        @num += chunk.size
      }
    end
  end
end

[repeatedly]

I forgot to send a patch...

@mururu Could you check this patch?

[mururu]

LGTM.
And I confirmed that I can reproduce the bug using the given conf (but I used in_dummy) and it disappears on this patch.