[Video Reader] add caching and retry limit for bad video tolerance

[zhreshold]

Add cached frame in case some video is corrupted and in the previous version it's silently handled by rewinding to nearest "good" frame, this will increase the chance of infinite rewinding if the video is completely corrupted.

So if a corrupted video is read by VideoReader with some bad frames(e.g., 200-300 out of 300 frames), retrieving vr[200] to `vr[299] will return vr[199] instead.

If you don't care about duplicate frames in this case, no worries.
If you want to prevent this from happening frequently(not ideal for training), then set the hard limit and catch the error on the fly to detect bad videos:

from decord import DECORDLimitReachedError
# accept at most 10 corrupted and recovered frames:
for video in videos:
    vr = VideoReader(video, fault_tol=10)
    # or percentage, e.g. 10% bad frames allowed
    # vr = VideoReader(video, fault_tol=0.1)
    try:
         # do what ever, read batch or loop
    except DECORDLimitReachedError:
        # handle bad video
    except DECORDError:
        # handle other types of error raised from decord