Support deep compression for embedding network with constant width

[ssrokyz]

Plz, support the deep compression feature with potentials with embedding network neurons of constant width.
Currently, multiple hidden-layered potentials with n+1=n*2 size only are supported.
In those cases, layers with layers more than 6 require too much memory, so fail to run.

If the hidden-layer size with n+1=n can be compressed, which utilizes skip connection too, it would be better to make deeper potentials that show better accuracy in some cases.

[ssrokyz]

Related previous question about the supported hidden-layer size of deep compression
#1370 (comment)

[njzjz]

ok, I'll mark it as a feature request.