Async Control Flow Refactor

[MichealReed]

• Refactored key GPU operations (createContext, toCPU, dispatchKernel, and createKernel) to return std::future instead of blocking with busy-wait loops.
• Replaced inline lambda callbacks with free‑standing functions that use promise/future for asynchronous completion.
• Updated createKernel (and its templated overload) to report shader module compilation asynchronously (instead of synchronously busy‑waiting), and dispatchKernel now returns a future that is fulfilled when the kernel dispatch is complete.
• Replaced wait with WaitForFuture(ctx.instance, future) as a template to return future objects after webgpu processing is complete.

Example usage:

  // Create kernel asynchronously
  std::future<Kernel> kernelFuture = createKernel(ctx, code, bindings, totalWorkgroups);
  Kernel kernel = waitForFuture(ctx.instance, kernelFuture);

  // Dispatch the kernel and wait for completion asynchronously
  std::future<void> dispatchFuture = dispatchKernel(ctx, kernel);
  waitForFuture(ctx.instance, dispatchFuture);

Why This Is Better:
• No blocking loops or synchronous waits – all async work is chained via futures, letting the event loop run naturally.
• Reduced risk of reentrancy issues and resource errors due to improper lifetime management.
• Cleaner separation between asynchronous callbacks and higher‐level logic, making the code easier to maintain and integrate (especially in WebAssembly/Emscripten contexts).

[MichealReed]

Only breaking change with latest commit is dispatchKernel and toCPU no longer take a promise, toCPUAsync and dispatchKernelAsync should be used instead to get futures. Manually waiting should no longer be necessary with the sync call though, even in a emscripten context.

[MichealReed]

Added an optional param with this refactor to set a offset for the toCPU buffer read. This is tested in the new test/test_gpu.cpp. I think the toCPU with copydata is now redundant. Will leave it alone if you prefer to keep it.