Fixes grouped Conv1d/Conv2d init to use in_channels // groups

[CodeWithMoin]

Conv1d and Conv2d computed the init bound using the full in_channels, while the weight's last dimension is in_channels // groups. For groups > 1, this made uniform init range too narrow (i.e 0.71x at 2 groups, 0.5x at 4, and 0.35x at 8). This PR fixes #4536 by using fan_in = (in_channels // groups) * kernel_size, that matches PyTorch's grouped-conv default. For the Conv3d and transposed convs, we don't have groups option, so they are unaffected. I've created a new test that checks init bound for groups 1, 2 and 4 on both layers. This test fails before the fix, and passes after the PR fix, In fact the full test_nn.py passes.

This changes only the initial random weights, so loaded checkpoints are unaffected.

Before (64 in/out channels, kernel 3, seed 0):

layer    g       weight shape fan_in   1/sqrt   max|w|  ratio
Conv1d   2        (64, 3, 32)     96   0.1021   0.0722  0.707
Conv1d   4        (64, 3, 16)     48   0.1443   0.0721  0.500
Conv2d   4     (64, 3, 3, 16)    144   0.0833   0.0417  0.500

After: ratio 1.000 for every group count, both layers.

Tests: test_nn 74 OK · test_conv 20 OK · test_conv_transpose 10 OK · test_init 11 OK

• ☑️ I understand it is strictly prohibited to use AI to write PR description
• AI usage disclosure: I've understood the init math, checked the test results then used assistance for code and new test.