Add FP16 supports

[taehoonlee]

TensorNets lacks of FP16 supports. For example, the following should work:

inputs = tf.placeholder(tf.float16, [None, 224, 224, 3])
model = nets.ResNet50(inputs)

[taehoonlee]

	float16 support	bonding with YOLOv2
ResNet50	>=1.5.0	O
ResNet101	>=1.5.0	O
ResNet152	>=1.5.0	O
ResNet50v2	>=1.5.0	O
ResNet101v2	>=1.5.0	O
ResNet152v2	>=1.5.0	O
ResNet200v2	>=1.5.0	O
ResNeXt50c32	>=1.5.0	O
ResNeXt101c32	>=1.5.0	O
ResNeXt101c64	>=1.5.0	O
WideResNet50	>=1.5.0	O
Inception1	>=1.1.0	O
Inception2	>=1.5.0	O
Inception3	>=1.5.0	X
Inception4	>=1.5.0	X
InceptionResNet2	>=1.5.0	X
NASNetAlarge	>=1.5.0	O
NASNetAmobile	>=1.5.0	O
VGG16	>=1.1.0	O
VGG19	>=1.1.0	O
DenseNet121	>=1.5.0	O
DenseNet169	>=1.5.0	O
DenseNet201	>=1.5.0	O
MobileNet25	>=1.5.0	O
MobileNet50	>=1.5.0	O
MobileNet75	>=1.5.0	O
MobileNet100	>=1.5.0	O
SqueezeNet	>=1.1.0	X
Darknet19	>=1.5.0	O

[taehoonlee]

The performance issue of `pretrained_initializer` with float16 should be investigated (see the workaround).