This PR is related to #69.

As mentioned in the readme (Finetuning & Inference, VQA, 4.inference section), we have prepared 2 types of inference for evaluation after fine-tuning. However, in the validation during fine-tuning, the user can only use all-candidate validation, which may be much slower. Now we have added a new option named val_inference_type in the fine-tuning scripts of VQA (see Line 61 in run_scripts/vqa/train_vqa_base_distributed.sh and run_scripts/vqa/train_vqa_distributed.sh). This option can be set as allcand (by default) or beamsearch, which stand for all-candidate and beam-search evaluation, respectively. Compared to the default all-candidate validation used in previous commits, switching to beam-search validation will be much faster (with around 0.5-0.6 validation score degradation compared with all-candidate validation).