Fix PESQ aborting a whole batch on one unscorable sample

[Kayvan-Zahiri]

Fixes #3304.

One unscorable sample takes down the whole batch. pesq and pesq_batch both
default to on_error=PesqError.RAISE_EXCEPTION and torchmetrics never passes
on_error, so what happens depends on n_processes:

• n_processes=1 (the default): pesq() raises NoUtterancesError and it
  propagates out of update(). The _filter_error_msg guard added in Ignore the NoUtterancesError when calculating pesq for a batch #2753 is
  unreachable on this path.
• n_processes != 1: pesq_batch() catches worker exceptions and returns them
  in the result list, so the guard is reached, but it drops the failed
  entry. A 3-sample batch silently returns 2 values.

Measured on a 3-sample batch with target[1] silent:

functional,   n_processes=1  -> raises NoUtterancesError
functional,   n_processes=2  -> shape (2,)  values [2.5199, 2.1842]   <- one sample vanished
class metric, n_processes=2  -> compute()=2.3521  total=2             <- 3 in, 2 counted

After:

functional,   n_processes=1  -> shape (3,)  values [2.5199, nan, 2.1842]
functional,   n_processes=2  -> shape (3,)  values [2.5199, nan, 2.1842]
class metric, either         -> compute()=2.3521  total=2

Change

Pass on_error=PesqError.RETURN_VALUES at all three call sites and map every
failure report onto nan, keeping one entry per input sample. Failures arrive
two ways and both are handled: negative error codes (-1..-7, which cannot
collide with a real score since valid PESQ is >= -0.5) and the exception
objects pesq_batch collects from its workers.

The class metric excludes nan samples from both the running sum and total,
so a degenerate sample does not poison an epoch's average. That matches what the
multiprocessing path already did by dropping them, so it is not a new policy.

Scores for scorable samples are unchanged: RETURN_VALUES and
RAISE_EXCEPTION return the same value on success (verified,
2.158228635787964 both ways).

Also in this PR

#2753 replaced pesq_val.reshape(preds.shape[:-1]) with
reshape(len(pesq_val)), so a (2, 3, 2100) input returned (6,) instead of
(2, 3), contradicting the documented (...,) shape. Restored, which is safe
now that no entries are dropped. This is a behaviour change for ndim > 2
inputs relative to 1.4.x-1.9.0. Happy to split it into its own PR if you would
rather keep this one minimal.

Tests

Four cases in tests/unittests/audio/test_pesq.py, all failing on main:

test_degenerate_sample_does_not_abort_batch[1]  NoUtterancesError
test_degenerate_sample_does_not_abort_batch[2]  assert torch.Size([2]) == (3,)
test_degenerate_sample_returns_nan_for_single_sample  NoUtterancesError
test_multidim_input_keeps_shape                 assert torch.Size([6]) == (2, 3)

The two different failure modes for [1] and [2] are the two paths above.

tests/unittests/audio/test_pesq.py: 20 -> 24 passed, same 6 skipped and 3
xfailed. Doctests pass. Wider audio dir: 88 passed. ruff, format and mypy clean.