Co-Authored-By: Ross Barnowski <rossbar@berkeley.edu>

This makes the C-API failed to load import error link to a new
troubleshooting document in the user documentation. It links
to the devdocs directly, since it should be possible to update the
information at any time (e.g. when a new setup starts making issues).
As these are not NumPy issues, their occurance is not tied to the
NumPy release process.

REV: Reverts side-effect changes to casting

This mostly reverts commit 6b954aa47bcce17cd7497d41f818950f847fbd8e
to ensure only the longdouble floating point numbers are changed, for
which this was a bugfix was relevant.
All other numbers should be unchanged (and were faster with the change),
but revert since this is a bugfix release.

Mainly the string->boolean conversions where accidentally changed to
arguably more correct behaviour. But this should not have happend
accidentally, especially in a bug-fix branch.
Another corner case change is that string->datetime casts within
hypothetical C-extensions doing manual C-level casting was fixed to be in
line with normal string -> datetime casts (the cast was inconsistently
defined).

This revert is only applied as a backport. The second change will remain
included in the next non-bugfix release while the first is fixed in
a more specific in cleaner manner.

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.

Assert that an invalid value (2**n-1 for n = 8, 16, 32, 64) has not
been passed to the Lemire function.

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.

BUG: Alpha parameter must be 1D in `generator.dirichlet`