[BugFix]Ensure that bf16 arrays are created as expected

[sisleyli]

If we use the bfloat16 type of ml_dtypes to create an array of tvm.ndarray, the result of the created tvm.ndarray will be wrong.

x = np.random.uniform(low=0.3, high=13.3, size=shape).astype(ml_dtypes.bfloat16)
# The values of x_ndarray and x are inconsistent!
x_ndarray = tvm.nd.array(x)

I understand that the reason for this error is that we have not considered that the input passed in is ml_dtypes.bfloat16. The np_float2np_bf16 function that has been used in the past (this function will convert the fp32 type numpy array into a uint16 numpy array)

[sisleyli]

Hi @masahi @Hzfengsy , could you please help me review this PR? Thank you!

[sisleyli]

@tvm-bot rerun