Reduce dnn top memory consumption

[dkurt]

• Replace double precision weights introduced at Remove setBatchNorm and setScale methods #10821 to fusion multipliers.
• Avoid Caffe weights copying during network upgrade.
• Release weights memory after parsing protobuf layer (Caffe and TensorFlow importers).

Maximum resident size for a single image forward pass according to /usr/bin/time --verbose:

(Importer) Network	Weights (MB)	output/internal blobs (MB)	Import + forward (MB) current	Import + forward (MB) PR	X-factor (PR vs. current)
(Caffe) AlexNet	233	3.4	726	398	0.5482093664
(Caffe) GoogLeNet	52	14.9	169	81.7	0.4834319527
(Caffe) ResNet-50	98	12.6	404	221.7	0.5487623762
(Caffe) SqueezeNet v1.1	4.8	8.35	45	34.7	0.7711111111
(Caffe) MobileNet-SSD	23	13	118	73.1	0.6194915254
(TF) MobileNet-SSD	28	13.55	136	82.4	0.6058823529
(TF) Inception 5h	52	15.47	161	113.4	0.7043478261
(Torch) OpenFace	31	1.92	76	72.8	0.9578947368
(Torch) Enet @ 512x512	3.1	65	88	86.1	0.9784090909
(Darknet) YOLOv3	237	54.7	1038	553.7	0.5334296724

allow_multiple_commits=1

[alalek]

Static analyzers dislikes this (NULL pointer dereference):

• add pointer check before (CV_Assert(ptr != NULL))
• or revert pointer to reference (but this check should be added in caller).

[dkurt]

@alalek, It was just to reduce number of modified lines. So I replaced all the v0_layer_connection. to v0_layer_connection->.

[alalek]

The problem is the same: needed an assertion check or non-NULL guarantee before using of "->".

[dkurt]

@alalek, thanks, got it! Added it here and at UpgradeV1Net and UpgradeV1LayerParameter.