Add BF16 convolution support to the XNNPACK backend

[JakeStevens]

Implement native BF16 convolution support in the XNNPACK backend.

Today BF16 convolution is intentionally excluded from XNNPACK delegation and falls back to portable kernels (see #23021).

Acceptance criteria:

• Support BF16 Conv1d and Conv2d delegation.
• Preserve correct tensor and weight datatypes during serialization.
• Add numerical tests covering delegation and execution.
• Remove or narrow the BF16 convolution fallback introduced by Keep unsupported BF16 and dynamic reshapes out of XNNPACK #23021.