wgpu: Use write_buffer_with rather than write_buffer

[waywardmonkeys]

This can be a performance improvement due to reduced data copying.

From the updated (post 0.20) docs:

    /// Write to a buffer via a directly mapped staging buffer.
    ///
    /// Return a [`QueueWriteBufferView`] which, when dropped, schedules a copy
    /// of its contents into `buffer` at `offset`. The returned view
    /// dereferences to a `size`-byte long `&mut [u8]`, in which you should
    /// store the data you would like written to `buffer`.
    ///
    /// This method may perform transfers faster than [`Queue::write_buffer`],
    /// because the returned [`QueueWriteBufferView`] is actually the staging
    /// buffer for the write, mapped into the caller's address space. Writing
    /// your data directly into this staging buffer avoids the temporary
    /// CPU-side buffer needed by `write_buffer`.
    ///
    /// Reading from the returned view is slow, and will not yield the current
    /// contents of `buffer`.
    ///
    /// Note that dropping the [`QueueWriteBufferView`] does *not* submit the
    /// transfer to the GPU immediately. The transfer begins only on the next
    /// call to [`Queue::submit`] after the view is dropped. To get a set of
    /// scheduled transfers started immediately, it's fine to call `submit` with
    /// no command buffers at all:

[waywardmonkeys]

(This is partly a reminder to myself. I planned to do this and keep forgetting.)

[waywardmonkeys]

Not sure if this actually an improvement, need to ask upstream.