Add bias keyword argument to cov - #691

* ENH: expose correction and weights parameters in cov

Resolves #688. Adds `axis`, `correction`, `frequency_weights`, and
`weights` to `cov`, giving users control over the degrees-of-freedom
correction and the observation-axis / weighted variants that
`numpy.cov` and `torch.cov` already support.

Naming follows array-api conventions (`axis`, `correction`) rather
than numpy's (`rowvar`, `bias`, `ddof`); the docstring includes a
one-to-one mapping. The delegation moves observations to the last
axis via `xp.moveaxis`, collapsing `rowvar` out of the backend
dispatch — only `ddof` vs `correction` differs between branches.

Dask's native `cov` forces `.compute()` on a lazy scalar when any
weights are given, so weighted dask inputs fall through to the
generic implementation, which is fully lazy.

* MNT: drop device= in cov weights

* STY: formatter

* TST: add bias tests from #691

* Update _funcs.py

Co-authored-by: Quentin Barthélemy <q.barthelemy@gmail.com>

* Update _delegation.py

Co-authored-by: Quentin Barthélemy <q.barthelemy@gmail.com>

* MNT: rename weights params to fweights/aweights

* ENH: validate weights shape in cov

* MNT: address lucascolley review

* MNT: move weights validation to generic cov

* Update _funcs.py

Co-authored-by: Quentin Barthélemy <q.barthelemy@gmail.com>

* DOC: explain non-integer correction use cases in cov

Addresses review feedback (kgryte, betatim) that the motivation for
allowing non-integer correction was not obvious from the docstring:
weighted unbiased correction and autocorrelated data both require
fractional values.

* TST: cover weight validation error paths in cov

Adds tests for the 1-D shape and length checks in the generic cov
path. Raises the diff coverage for this PR from 93.33% to 100%.

* Preserve torch autograd in the batched cov path

The generic `cov` implementation called `xp.asarray(m)` on its input. For
torch this detaches gradients and mutates the caller's tensor in place, so
`cov` on a batched tensor (ndim > 2, which routes to the generic path) with
`requires_grad=True` returned a detached result and silently zeroed the
input's grad.

The call is unnecessary: the delegation layer already guarantees `m` is an
array (it calls `array_namespace(m)` and reads `m.ndim`). Drop it, and add
a torch autograd regression test.

* MNT: address cov review feedback

* TYP: clarify array cast in cov warning test

* lint

* Apply suggestions from code review

Co-authored-by: Lucas Colley <lucas.colley8@gmail.com>

* MNT: address cov review comments

- link numpy/torch functions with intersphinx in the cov docstring
  (torch added to the intersphinx mapping)
- add canonical examples for correction, fweights and aweights
- explain the int(correction) cast: torch.cov rejects integer-valued
  floats such as 1.0 at runtime, so typing.cast is not enough
- simplify the integer-correction check to float(correction).is_integer()
- comment that the NaN shape checks account for Dask reporting unknown
  dimensions as NaN instead of None

---------

Co-authored-by: Quentin Barthélemy <q.barthelemy@gmail.com>
Co-authored-by: Lucas Colley <lucas.colley8@gmail.com>