Fix JVPs of power, divmod, and slice_update with partly traced inputs

[qflen]

Fixes #3634

Same class as #3627/#3629 (fixed by #3633):

Primitive::jvp implementations that break the packed-tangents convention (tangents[i] is the tangent of input argnums[i]; a jvp returns one tangent per output).

jvp	bug	fix
Power	delegated to vjp, which scales every partial by cotangents[0], so with both inputs traced both partials used the first input's tangent and the second tangent was ignored	one vjp call per traced input with its own tangent as the cotangent, summed
DivMod	returned one tangent for a two-output primitive, so the second output never entered the tangent map and its consumer was invoked with empty argnums/tangents — segfault	return one zero tangent per output (divmod stays non-differentiable, matching its vjp)
DynamicSliceUpdate	ignored argnums and read tangents[0]/tangents[1] unconditionally — OOB segfault with one traced input, mispaired tangents for argnums = {0, 2}	untraced inputs get zero tangents; traced start indices throw, matching the vjp

Before

_, (j,) = mx.jvp(lambda a, b: a**b, [mx.array(2.0), mx.array(3.0)],
                 [mx.array(0.0), mx.array(1.0)])
# j = 0.0, silently wrong (vjp and finite differences give ln(2)·8 ≈ 5.5452)

The divmod and slice_update repros in #3634 segfaulted.

After

j = 5.5452, jvp matches vjp and finite differences, and all repros pass.

Added a regression test covering tangent pairing for power, per-output tangents for divmod (including a consumer of the second output), and dynamic slice_update with each subset of traced inputs. Each case fails on unfixed main (wrong value for power, segfault for the other two).

Checklist

• I have read the CONTRIBUTING document
• I have run pre-commit run --all-files to format my code / installed pre-commit prior to committing changes
• I have added tests that prove my fix is effective or that my feature works
• I have updated the necessary documentation (if needed)

[zcbenz]

Thanks!