Resampling appears to use resample-by-matrix path

[inducer]

@matthiasdiener showed this GPU profile:

Resampling kernels (there are two types: "resample-by-picking" and "resample-by-matrix", cf here) take up much of our time. I don't see a good reason why resample-by-matrix should be used right now. (given that we're not yet using overintegration)

See also #62.