Add fuse layers for conv+affine+relu and conv+relu

[facug91]

This PR extends the fuse_layers method to also fuse convolutional layers followed by relu, and convolutional layers follower by affine followed by relu.

I tested it with examples/dnn_mmod_face_detection_ex.cpp and got exactly the same results on the images shown.

Here are also some tests for different neural networks from https://github.com/dlibml/dnn:

Name	master (w/o fusion)	master (w/ fusion)	this PR (w/o fusion)	this PR (w/ fusion)	comparative1	comparative2
alexnet	9.53	9.63	9.55	9.45	1.84%	1.10%
sqznet1.0	6.53	6.00	6.40	4.47	25.60%	30.25%
sqznet1.1	4.41	3.90	4.41	3.01	22.86%	31.76%
vggnet11	37.00	34.46	36.76	32.77	4.91%	10.87%
vggnet13	46.94	43.90	47.31	41.06	6.47%	13.21%
vggnet16	58.80	54.87	58.01	52.29	4.71%	9.86%
vggnet19	68.32	65.11	69.29	61.89	4.95%	10.69%
googlenet	9.95	9.54	10.09	8.53	10.61%	15.50%
resnet18	7.66	7.39	7.68	6.86	7.23%	10.64%
resnet34	13.90	13.67	13.90	12.77	6.57%	8.10%
resnet50	22.41	20.11	22.41	18.89	6.07%	15.71%
resnet101	39.15	35.72	38.39	33.73	5.56%	12.13%
resnet152	54.98	50.76	55.03	48.44	4.57%	11.98%
darknet19	13.55	12.43	13.54	12.53	-0.79%	7.51%
darknet53	32.70	30.87	32.36	30.63	0.80%	5.36%
darknet53csp	26.47	23.36	26.60	23.48	-0.53%	11.73%
densenet121	21.57	20.33	21.57	20.12	1.03%	6.71%
densenet169	28.50	27.13	28.43	25.66	5.43%	9.76%
densenet201	38.60	37.17	38.52	35.65	4.10%	7.45%
densenet265	55.85	54.43	55.81	52.72	3.13%	5.53%
densenet161	49.80	47.90	49.71	45.74	4.52%	7.99%
vovnet19s	6.84	6.39	6.85	5.05	20.88%	26.23%
vovnet19	14.25	13.85	14.18	11.57	16.47%	18.41%
vovnet27s	8.31	7.84	8.31	6.25	20.31%	24.81%
vovnet27	18.32	17.67	18.32	15.28	13.54%	16.61%
vovnet39	23.69	23.30	23.53	20.21	13.23%	14.11%
vovnet57	31.00	30.81	31.02	27.41	11.03%	11.63%
vovnet99	56.88	56.44	56.31	51.45	8.84%	8.64%
repvgg_a0	6.47	5.67	6.38	4.96	12.52%	22.22%
repvgg_a1	8.01	8.06	8.04	7.23	10.30%	10.14%
repvgg_a2	19.50	19.46	19.35	18.01	7.49%	6.95%
repvgg_b0	10.16	10.17	10.04	8.88	12.61%	11.54%
repvgg_b1	42.13	42.18	42.12	39.34	6.72%	6.60%
repvgg_b2	59.09	59.01	58.67	55.87	5.32%	4.78%
repvgg_b3	82.28	82.40	82.66	79.12	3.97%	4.28%

comparative1: measures how much faster the model run with the new fuse_layers compared to the current fuse_layers.
comparative2: measures how much faster the model run with the new fuse_layers compared to not using fuse_layers.

I left the measurements before the fusion to be sure that the different runs used the same convolutional algorithms in cuDNN. Every time they differed by more than 2% up or down, I ran those tests again.

[facug91]

test_fuse_layers is failing, so I must have done something wrong. I'll review this before removing the Draft. Maybe @arrufat can help me out here please? This is the PR I was talking about before 😅

[arrufat]

I'm traveling these days, so it might take some take to have a look. I'll try during this week, though.

I also thought about fusing the relu layer, so it's nice so see it here.

[facug91]

I thought it was test_fuse_layers the test that was failing, but no, it was something else. When I looked for the line in the test file, it was a test of a normal convolution. I had forgotten to update the copy constructor and the assignment operator, and that's why it was failing. The tests on the CPU pass correctly now.
Anyway, I noticed that the tests on gpu are not passing. But it is not a problem of this PR, in master they are not passing either.

[facug91]

One other thing, while reviewing this problem, I was looking for the documentation of "disable_duplicative_biases" function and I couldn't find it. Turns out it was left in layers_abstract.h file, so I took the opportunity to made a commit moving it to "visitors_abstract.h" right here.

[davisking]

disable_duplicative_biases

Oh yeah, thanks. That fixes the links in the main docs too :)

[davisking]

Huh yeah, the GPU version of, layer_norm_ is failing since it's giving out bad derivatives on master. We need a CI test that runs with a GPU :|

Anyway, yeah, looks like your PR is good :)

[arrufat]

Huh yeah, the GPU version of, layer_norm_ is failing since it's giving out bad derivatives on master. We need a CI test that runs with a GPU :|

Anyway, yeah, looks like your PR is good :)

Oh, I need to check what's going on there