fix: guard group<=0 in ConvolutionDepthWise(::1D)::load_param to avoid FPE divide-by-zero (#6911)

[FluxMind-O]

Summary

Fixes #6911 — FPE divide-by-zero in ConvolutionDepthWise::load_param when a malformed param file provides group == 0.

Problem

When group == 0, the existing guard num_output % group != 0 itself performs an integer division by zero, triggering SIGFPE (FPE divide-by-zero) before the rejection path is reached. The CPU and 1D variants of ConvolutionDepthWise::load_param were missing the group <= 0 check that the Vulkan int8 path already had (convolutiondepthwise_vulkan.cpp:597).

Fix

Add a group <= 0 short-circuit guard before the modulo in both affected load_param implementations, returning -100 (the existing "reject invalid group" code) without performing the division:

if (group <= 0 || num_output % group != 0)
{
    // reject invalid group
    return -100;
}

• src/layer/convolutiondepthwise.cpp — CPU ConvolutionDepthWise::load_param
• src/layer/convolutiondepthwise1d.cpp — ConvolutionDepthWise1D::load_param

The Vulkan variant calls the CPU load_param, so it is covered transitively. group < 0 is rejected too (defensive), matching the spirit of the existing guard. Normal-path behavior is unchanged: valid group > 0 with num_output % group == 0 still returns 0.

Regression tests

• tests/test_convolutiondepthwise_invalid.cpp — asserts load_param returns -100 for group = 0, -1, -8.
• tests/test_convolutiondepthwise1d_invalid.cpp — same for the 1D variant.

Verification

• Before fix: test_convolutiondepthwise_invalid aborts with Floating point exception (core dumped) (exit 136 / SIGFPE), reproducing bug: FPE (divide-by-zero) in ConvolutionDepthWise::load_param (src/layer/convolutiondepthwise.cpp:46) #6911.
• After fix: both new tests exit 0.
• Existing test_convolutiondepthwise, test_convolutiondepthwise_1, test_convolutiondepthwise_oom, test_convolutiondepthwise1d all still pass — no regression on the normal path.

Checklist

• one clear bug -> one local fix -> one regression test
• no change to normal-path behavior
• no architecture / API change

[tencent-adm]

Thank you for your submission, we really appreciate it. Like many open source projects, we ask that you sign our Contributor License Agreement before we can accept your contribution.
_{You have signed the CLA already but the status is still pending? Let us recheck it.}

[FluxMind-O]

Closing as duplicate — my apologies to the maintainers.

After opening this I found that #6912 by @beilzx already fixes #6911 (same group == 0 guard on the CPU + 1D load_param paths, CLA signed, open for review). There is no value in keeping a second PR for the same one-line guard, so I'm withdrawing this one to avoid wasting reviewer time.

Thanks @beilzx for the prompt fix. I'll go review #6912 instead.