I found the BSmooth (http://www.ncbi.nlm.nih.gov/pubmed/23034175) paper provides a justification for the use of smoothing:

This has led most WGBS studies to employ a high coverage design since even 30× coverage yields standard errors as large as 0.09. However, various authors have noted that methylation levels are strongly correlated across the genome [24,25]. Furthermore, functionally relevant findings are generally associated with genomic regions rather than single CpGs, either CpG islands [26], CpG island shores [27], genomic blocks [1], or generic 2 kb regions [3].

They then concluded the following:

Using this method [BSmooth] on data with 4× coverage, we achieved precision comparable to deeper coverage without smoothing.

So my guess is that one answer could be that smoothing/windows allowed lower coverage sequencing through still having low standard errors associated with the (average/smoothed) DNA methylation level. This is of course at the cost of resolution in resolving individual CpGs.

My guess is that they once had a dataset with either low coverage or a lot of noise. The sliding window would allow you to handle that and still assign values to focal regions/points. There's no other good reason that I know of to do this and it's not something I would personally do by default.