np.bool was removed in NumPy 1.24.

This always produced invalid results (though they were consistent
between Numba and NumPy) but now this fails in NumPy 1.24 with an
exception:

```
TypeError: The `dtype` and `signature` arguments to ufuncs only select
the general DType and not details such as the byte order or time unit.
You can avoid this error by using the scalar types `np.float64` or the
dtype string notation.
```

Note that the exception message is misleading, and using the dtype
string notation does not provide a workaround.

I believe this was written in error and should always have been float16.

The modified regex matches the existing message produced by NumPy <
1.24, and the new improved message in 1.24.

Setting an array element with a sequence is removed in NumPy 1.24.

This is so that a proper solution can be implemented.

This reverts commit 08416ad.

`init_ufunc_dispatch()` should set an exception if it's going to fail.
Without setting an exception, an unrecognized ufunc method (e.g.
`resolve_dtypes()` in NumPy 1.24) results in:

```
SystemError: initialization of _internal failed without raising an
exception
```

This commit modifies it to set an exception notifying the user of the
error. Now when an unrecognized ufunc method is encountered, the
traceback ends in:

```
    from numba.np.ufunc import _internal
RuntimeError: Unexpected ufunc method resolve_dtypes()
```

The Windows build did not like a variable-length stack array.

The API version has long since been greater than 0x7 / 0x8 for any
supported NumPy.

This reverts commit 22980fe.

This test only checked for a plain match when comparing outputs.
However, in some cases a reconstruction check can be necessary, as in
`test_linalg_svd`.

PR numba#8691 feedback.