Add naive random masking strategy for masking timeseries

[geek-yang]

As a follow-up to PR #554, we can add a naive fully random masking strategy for masking timeseries. For instance, there are two simple ways for it:

• Simply flatten the input data (from [timesteps, channels] to [timesteps * channels]) and mask them randomly
• Loop through all channels and treat them individually (mask each channel separately and concatenate them)

These two strategies seem to be the approaches taken by "Lime-for-time" implementation. Given that this is a low-hanging fruit and we have no idea whether the smart masking works better than this simple masking or not, it is worth-while to add this.

[cwmeijer]

This is already possible by setting the parameters correctly. This issue should be resolved by documenting this well in the right places and probably by using this naive approach as a default behavior.