Shared tensor plumbing behind de_sample.nsbi_de_maf and de_sample.nsbi_de_nsf

Takes the single conditioning row, replicates it to n rows, draws a standard-normal base sample, and inverts the flow – all on the net's device (see net_device()), then brings the draws back to CPU as a plain R matrix. inverse_fn is the per-flow inverse – maf_inverse() or nsf_inverse(). The MDN has no inverse to share here; it samples its mixture directly.

Usage

de_sample_flow(de, x, n, inverse_fn)