Fix Net.forward_backward_all losing entries when outputs and inputs count differ

[Chessing234]

Bug

`Net.forward_backward_all` (attached from `_Net_forward_backward_all` in `python/caffe/pycaffe.py`) silently leaves some entries as un-converted Python lists, potentially still padded with zero rows, whenever the net has a different number of outputs+extra blobs than inputs+extra diffs.

Root cause

After collecting per-batch forward and backward results into two separate dicts (one keyed by `self.outputs ∪ blobs`, the other by `self.inputs ∪ diffs`), the function wraps up with:

```python

Package in ndarray.

for out, diff in zip(all_outs, all_diffs):
all_outs[out] = np.asarray(all_outs[out])
all_diffs[diff] = np.asarray(all_diffs[diff])

Discard padding at the end and package in ndarray.

pad = len(six.next(six.itervalues(all_outs))) - len(six.next(six.itervalues(kwargs)))
if pad:
for out, diff in zip(all_outs, all_diffs):
all_outs[out] = all_outs[out][:-pad]
all_diffs[diff] = all_diffs[diff][:-pad]
```

`all_outs` and `all_diffs` are independent dicts with independent key sets. `zip(all_outs, all_diffs)` stops at the shorter of the two, so any excess entries in the longer dict:

• are never turned into ndarrays by `np.asarray` — they stay as Python lists of numpy arrays, and
• are never trimmed to the real batch size, so they keep the zero-padding appended at the end of the last partial batch in `_Net_batch`.

This bites every net with a different number of outputs than inputs, which is extremely common (e.g. one-input-multiple-heads classification+regression nets). Only the `len(all_outs) == len(all_diffs)` case (1-input, 1-output, no extras) happens to work.

Fix

Walk each dict independently instead of zipping them together:

```diff
# Package in ndarray.

• for out, diff in zip(all_outs, all_diffs):

•    all_outs[out] = np.asarray(all_outs[out])
  

•    all_diffs[diff] = np.asarray(all_diffs[diff])
  

• for d in (all_outs, all_diffs):

•    for k in d:
  

•        d[k] = np.asarray(d[k])
  

  Discard padding at the end and package in ndarray.

  pad = len(six.next(six.itervalues(all_outs))) - len(six.next(six.itervalues(kwargs)))
  if pad:

•    for out, diff in zip(all_outs, all_diffs):
  

•        all_outs[out] = all_outs[out][:-pad]
  

•        all_diffs[diff] = all_diffs[diff][:-pad]
  

•    for d in (all_outs, all_diffs):
  

•        for k in d:
  

•            d[k] = d[k][:-pad]
  

```

No behavioral change for nets where `len(all_outs) == len(all_diffs)` — the prior `zip` already covered every entry in that case; every other shape is now correct too.