[MNT] SIGReg cleanup: dedup tests, backfill docstring, document weights (#1925)

* Remove duplicate TestSIGReg class

* LeJEPALoss distributed test mock consistency

TestLeJEPALoss.test_forward_gather_distributed_world_size_gt_one
was using the pre-1923 mock pattern that only patched
torch.distributed.all_reduce.
The assertion call_count == 3 passed only because
torch.distributed.nn.all_reduce is internally implemented via
torch.distributed.all_reduce, so the mock caught both APIs together.

Mock the two APIs separately and assert their counts individually
(c10d all_reduce: 1 call for num_samples_tensor,
nn all_reduce: 2 calls for cos_sum and sin_sum),
matching the pattern TestSIGReg uses after #1923.

* Backfill SIGReg class docstring

* Document trapezoidal weights symmetry exploit

Added an inline comment explaining the symmetry exploit,
the SIGReg integrand (cos_mean - phi)^2 + sin_mean^2 is even in t,
so the integral over [-t_max, t_max] equals twice the integral over [0, t_max].
Evaluating at only the non-negative knots and doubling the
trapezoidal weights gives the same result while halving the
number of knots to evaluate.

---------

Co-authored-by: fruet <90271657+gabrielfruet@users.noreply.github.com>