Fixes #16539

The implicit 0-padding that is done to small entropy inputs to make them the
size of the internal pool conflicts with the spawn keys, which start with an
appended 0.

In order to maintain stream compatibility with unspawned `SeedSequence`s, we
explicitly 0-pad short inputs out to the pool size only if the spawn key is
provided, and thus would trigger the bug. This should minimize the impact on
users that were not encountering the bug.

Broadcastable size with inputs does not produce an error when size produces
a smaller output array than the broadcast input shape. Patch checks that
the output shape matches the outer shape of the broadcast fo all inputs and
the size when given.

Refactor polynomial to be unsigned long array
Remove unused code
Fix md5 calculation on BE

Use the original loop order instead of an inverted order

closes #15394

BUG: add missing numpy/__init__.pxd to the wheel

* Add stick-breaking

* Add tests demonstrating slowness for beta and dirichlet generators for small alpha (and beta) values

* Remove the test for beta with small `a` and `b`

* Switch from standard to stick-breaking method whenever alpha.max() < 0.1

Co-authored-by: Warren Weckesser <warren.weckesser@gmail.com>
Co-authored-by: Eric Wieser <wieser.eric@gmail.com>
Co-authored-by: Sebastian Berg <sebastian@sipsolutions.net>

Add repeatability tests for when the range of the integers is `2**32`
(and `2**32 +/- 1` for good measure) with broadcasting.  The underlying
functions called by Generator.integers and random.randint when the
inputs are broadcast are different than when the inputs are scalars.

When the input to Generator.integers was 2**32, the value 2**32-1
was being passed as the `rng` argument to the 32-bit Lemire method,
but that method requires `rng` be strictly less then 2**32-1.

The fix was to handle 2**32-1 by calling next_uint32 directly.
This also works for the legacy code without changing the stream
of random integers from `randint`.

Closes gh-16066.