[webgpu] Bump version of Dawn to 991810463a (#25011)
### Description
This PR is a redo of the reverted PR
https://github.com/microsoft/onnxruntime/pull/24935.
- Upgrade version of Dawn to 991810463a.
- Comparing to previous commit c3999d7e3, it includes the fix to GCC13,
so that now Linux build passes.
- Includes previous merged patches:
- https://dawn-review.googlesource.com/c/dawn/+/242335
- https://dawn-review.googlesource.com/c/dawn/+/242634
- https://dawn-review.googlesource.com/c/dawn/+/243614
- Remaining patches:
- `dawn_destroy_buffer_on_destructor.patch`
**Allow WGPUBufferImpl class to destroy the buffer in the destructor**
In native implementation, wgpuBufferRelease will trigger the buffer
destroy (if refcount decreased to 0). But in emwgpu implementation, the
buffer destroy won't happen. This change adds a destructor to the buffer
class to destroy the buffer when the refcount is 0 for non-external
buffers.
- `dawn_force_enable_f16_nvidia_vulkan.patch`
**Force enable f16 support for NVIDIA Vulkan**
Dawn disabled f16 support for NVIDIA Vulkan by default because of
crashes in f16 CTS tests (crbug.com/tint/2164).
Since the crashes are limited to specific GPU models, we patched Dawn to
remove the restriction.
- `dawn_allow_non_uniform_for_subgroup_matrix_args.patch` (new)
**Revert change "[tint] Check uniformity for subgroup matrix builtin
arguments"**
The following change was made in upstream to force the subgroup matrix
builtin arguments to be uniform:
https://dawn-review.googlesource.com/c/dawn/+/236054
(https://issues.chromium.org/issues/403611487)
Since we use `subgroup_id` as the subgroup matrix builtin argument, we
have to revert this change to allow the subgroup matrix builtin
arguments to be non-uniform.
