Fast `to_homogeneous` conversion / RelGraphConv on heterogeneously sampled blocks

## 🚀 Feature
This issue is to discuss the usage of "fused heterogeneous" layers that work on sampled DGLBlocks which process the heterogeneous subgraphs as homogeneous (eg. `RelGraphConv`). There is a need to use a heterogeneous representation of a graph during the sampling level and then process the resulting blocks as a homogeneous graph to make use for accelerated methods that work on a such format.

## Motivation

We are interested in using homogeneous methods for processing heterogeneous graphs. Methods like RGCN can be accelerated by using their homogeneous versions like `RelGraphConv`. We want to keep the sampling, data processing logic clean and aligned with our methods implemented in PyG (where the preferred behaviour is to sample heterogeneously and the use the homogeneous processing on such sampled subgraphs).

## Alternatives

An alternative is to convert the graph before sampling. However, we are then sampling the graph in a homogeneous manner, which is incomparable to heterogeneous methods and often results in worse training results.

## Pitch

We would either want an accelerated `to_homogeneous` method (or its alternative) or a way to use homgeneous processing on heterogeneous, sampled DGLBlocks.

## Additional context
.