GeezerVison

SuiteSparse compares the byte size of an imported values buffer with the
number of entries, not with the bytes they need (GB_import.c sets
Ax_memsize_for_non_iso to nvals or vlen * vdim), so any buffer of at least
nvals bytes got through and was read past its end:

    Vector.ss.import_full(values=np.ones(3), size=5)

built a Vector whose last two values were heap garbage, and so did values
too short for a sub-array dtype, as in

    Vector.ss.import_sparse(size=3, indices=[0, 1], values=np.ones(2), dtype="FP64[2]")

Every ss import and pack (sparse, bitmap and full Vectors; csr, csc,
hypercsr, hypercsc, bitmapr, bitmapc, fullr and fullc Matrices, and so
from_dcsr and from_dcsc) now raises

    ValueError: values must have 5 or more elements of dtype FP64, not 3

before handing GraphBLAS the buffer. An iso import needs one element.

…lues

SuiteSparse:GraphBLAS 10.5.2 fixed the import size check that
check_values_size works around: GB_import.c compared the byte size of
the values with the number of entries. With check_values_size bypassed,
the five short imports tried (import_full, import_sparse of FP64[2],
import_bitmapr, and two import_csr) now raise InvalidValue with an empty
message on 10.5.2, where 10.5.1 built them from bytes past the end.
check_values_size still runs on every version: older ones need it, and
its ValueError says what is wrong. Its docstring and the comments of
the two test_ss_import_values_size tests now give the versions.
Comments only.

Matrix.from_dcsr([0, 1], [0, 1, 2], [0, 1], 1.5, "FP64[2]") gave
[[1.5, 1.5], [1.5, 1.5]] without SuiteSparse, and on SuiteSparse raised
"values must have 1 or more elements of dtype FP64[2], not 0": its
SuiteSparse branch imported the scalar as it was, 8 bytes where one
element is 16. It now broadcasts a scalar to one whole element and
imports it iso, as from_csr and from_csc do; the same for from_dcsc.

check_values_size also says what is wrong in two cases it misreported:
- A scalar that does not fill one element of a sub-array dtype, as in
  Vector.ss.import_full(values=1.5, size=3, dtype="FP64[2]",
  is_iso=True): "a scalar does not fill one element of dtype FP64[2], of
  shape (2,)" (it said "not 0").
- Index pointers that give more entries than there are indices, as in
  import_csr(nrows=1, ncols=3, indptr=[0, 2**62], col_indices=[0],
  values=[1.0]): "index pointers give 4611686018427387904 entries, more
  than the number of indices, 1" (it asked for that many values).
  GraphBLAS rejects such an indptr itself, with an InvalidValue and no
  message, so the check comes first.

import_fullr(values=1.5, nrows=2, ncols=2) said "a scalar does not fill
one element of dtype FP64", though 1.5 is a whole FP64 element and the
problem is that four are needed. Keep that message for a scalar shorter
than one element of a sub-array dtype. Found by review (r47abc).