Remove custom newSeqUninit if available in stdlib (#592)

* rename `newSeqUninit` in order to avoid conflicts with upstream

* use system `newSeqUninit` where applicable

In these cases we are only dealing with mem copyable POD types, so
it's fine to use the stdlib version.

* use regular `newSeq` in `toRawSeq`

The proc is deprecated anyway and the minor performance penalty should
be irrelevant.

* use system `newSeqUninit` if available

---------

Co-authored-by: ringabout <43030857+ringabout@users.noreply.github.com>

src/arraymancer/private/sequninit.nim

@@ -12,9 +12,9 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
+# from Nim https://github.com/nim-lang/Nim/pull/22739 on the stdlib provides a
+# `newSeqUninit` for types supporting `supportsCopyMem`
 when not declared(newSeqUninit):
-  # https://github.com/nim-lang/Nim/pull/22586#issuecomment-1698160304
-
   func newSeqUninit*[T](len: Natural): seq[T] {.inline.} =
     ## Creates an uninitialzed seq.
     ## Contrary to newSequnitialized in system.nim this works for any subtype T

src/arraymancer/tensor/exporting.nim

@@ -37,7 +37,7 @@ proc toRawSeq*[T](t:Tensor[T]): seq[T] {.noSideEffect, deprecated: "This proc ca
   # Due to forward declaration this proc must be declared
   # after "cpu" proc are declared in init_cuda
   when t is Tensor:
-    result = newSeqUninit[T](t.size)
+    result = newSeq[T](t.size)
     for i in 0 ..< t.size:
       when T is KnownSupportsCopyMem:
         result[i] = t.unsafe_raw_offset()[i]