Fix pytorch for chainer v6.0.0

[Fhrozen]

The corrections for pytorch with chainer v6.0.0 are included in this PR, to solve #760.

• Modified LMTest; this returned gpuid as a GPUId class, set self.device inside init
• Replaced extensions.snapshot_object for (custom) snapshot_object. V6.0.0 call first serialize and the use the savefun method (not sure why it does this), so the training will raise error due to the lack of serialize routine. While v5.0.0 is straight. Just copied the required code into the 'asr_utils' and replace the line in 'pytorch/asr/train' and 'pytorch/tts/train'.

I did not find any additional problem for pytorch with chainer v6.0.0 and lm_train, asr_train, asr_decode, and tts_train worked well.
I suppose #761 will no longer be required to merge.

[sw005320]

I found that it is appeared in the other parts (see https://github.com/espnet/espnet/search?q=extensions.snapshot_object&unscoped_q=extensions.snapshot_object).

[Fhrozen]

I just modified the last one (asr_mix) for pytorch. In case of chainer, this does not have any problem.

[sw005320]

Thanks, @Fhrozen!

[jnishi]

Thanks. LGTM!