Fix feature_matching loss weight applied twice (400x instead of 20x)

[taemincho]

Bug

In RAVE.training_step, entries added to loss_gen are already weighted at
construction time, e.g.:

loss_gen['feature_matching'] = self.weights['feature_matching'] * feature_matching_distance
loss_gen['adversarial'] = self.weights['adversarial'] * loss_adv

But the final accumulation loop re-applies the same weight via
self.weights.get(k, 1.):

loss_gen_value = 0.
for k, v in loss_gen.items():
    loss_gen_value += v * self.weights.get(k, 1.)

Any key in loss_gen whose name also exists in self.weights gets
double-weighted. With the default config, feature_matching's weight (20)
is effectively squared to 400x. adversarial (weight 1.0) is unaffected
numerically, and the spectral distance terms escape only because their
dict keys (multiband_spectral_distance, fullband_spectral_distance)
don't match self.weights's keys (multiband_audio_distance,
audio_distance).

This regression was introduced in 62a168a ("merge with last version +
normalization + bug fixes + input/output transforms"), which replaced the
previous, correct loss = sum(loss_gen.values(), 0) with the current loop,
without removing the pre-existing per-key weighting.

Fix

Since every entry in loss_gen is already weighted before insertion, the
final loop should just sum the values without re-weighting:

loss_gen_value = sum(loss_gen.values(), 0.)
loss_gen_value.backward()

This restores the original (pre-62a168a) behavior.