fix: ensure command encoder cleanup always runs

[jeffcrouse]

Summary

The Drop implementation for WGPUCommandEncoderImpl only ran cleanup when self.open was true, but wgpuCommandEncoderFinish set open=false before the encoder was released. This meant command_encoder_drop() was never called, causing memory to leak.

The Bug

// In wgpuCommandEncoderFinish (line 1401):
command_encoder.open.store(false, atomic::Ordering::SeqCst);

// In Drop (line 129):
if self.open.load(atomic::Ordering::SeqCst) && !thread::panicking() {
    context.command_encoder_drop(self.id);  // Never reached!
}

The Fix

Remove the open check from Drop - cleanup should always run when the encoder is released:

if !thread::panicking() {
    context.command_encoder_drop(self.id);
}

Test Results

Version	Memory at 0s	Memory at 180s	Behavior
Before fix	23 MB	33 MB	+0.5 MB every 10 seconds
After fix	21 MB	21 MB	Stable

Reproduction

Minimal C++ test that creates command encoders at 60fps:

while (true) {
    WGPUCommandEncoder encoder = wgpuDeviceCreateCommandEncoder(device, nullptr);
    WGPURenderPassEncoder pass = wgpuCommandEncoderBeginRenderPass(encoder, &passDesc);
    wgpuRenderPassEncoderEnd(pass);
    wgpuRenderPassEncoderRelease(pass);
    WGPUCommandBuffer cmdBuf = wgpuCommandEncoderFinish(encoder, nullptr);
    wgpuQueueSubmit(queue, 1, &cmdBuf);
    wgpuCommandBufferRelease(cmdBuf);
    wgpuCommandEncoderRelease(encoder);
    std::this_thread::sleep_for(std::chrono::milliseconds(16));
}

Fixes #541
Fixes #527

Notes

• The open flag still serves a purpose in wgpuCommandEncoderFinish to prevent double-finishing
• This fix ensures cleanup happens regardless of whether finish() was called
• Tested on macOS Sequoia / Apple M4 Max with wgpu-native v27

[cwfitzgerald]

Nice