A "sha256:<hex>" digest over the label array (dimensions and row-major
little-endian int32 values), the mask type and the effective legend.
Identical content always yields the same revision, so repeated imports of
the same prediction are idempotent.
Examples
m <- new_sg_mask(matrix(c(0L, 1L, 1L, 0L), 2))
sg_mask_revision(m)
#> [1] "sha256:aae04e92b620bf9e445573b0730a9f46997771628a7eac24eb723f0f32b6a2c9"