Add support for grouped 1D convolutions to the nn API (#1444)

* Fix the weight shape for grouped convolutions from the nn API.

* Add tests.

* Pre-commit formatting.

* Add input validation.

* Use integer division instead of casting.

* docs

* nit

---------

Co-authored-by: Awni Hannun <[email protected]>

python/tests/test_nn.py

@@ -650,6 +650,14 @@ class TestLayers(mlx_tests.MLXTestCase):
         c = nn.Conv1d(in_channels=C_in, out_channels=C_out, kernel_size=ks, bias=False)
         self.assertTrue("bias" not in c.parameters())
 
+        groups = C_in
+        c = nn.Conv1d(
+            in_channels=C_in, out_channels=C_out, kernel_size=ks, groups=groups
+        )
+        y = c(x)
+        self.assertEqual(c.weight.shape, (C_out, ks, C_in // groups))
+        self.assertEqual(y.shape, (N, L - ks + 1, C_out))
+
     def test_conv2d(self):
         x = mx.ones((4, 8, 8, 3))
         c = nn.Conv2d(3, 1, 8)