Add reflect, symmetric and edge padding

[utkuozbulak]

Added reflect, symmetric and edge padding options.

This option was requested by #400. PR #213 was sent to include the same functionality but still awaiting response from the author since Sep 20, 2017.

The new padding options also work the same way with how it is for the constant one.
If padding is an int: it pads all sides.
if padding is a tuple with two elements: it pads left/right and top/bottom respectively.
if padding is a tuple with four elements: it pads left, top, right, bottom borders respectively.
It also does not break or change the behavior of existing code.

In #213 @alykhantejani suggested that the function should be
pad(img, padding, padding_mode='constant', fill=0) instead of
pad(img, padding, fill=0, padding_mode='constant') because when the padding_mode is not constant, fill is irrelevant/not used. This is correct, but doing so might break existing code for some people so I did not change the positions.

Below are some tests:
RGB test image:

pad(img, 20, padding_mode = 'reflect')

pad(img, (0, 50), padding_mode = 'edge')

pad(img, (1, 10, 20, 30), padding_mode = 'symmetric')

Grayscale test image:

pad(img, (5,10,20,50), padding_mode = 'symmetric')

[fmassa]

This is awesome, thanks a lot!

Apart from some doc changes, I think it would be great if you could add some very simple tests for it.
Something simple, for example with a dummy image like

a = torch.zeros(10, 10, 3)
a[:, :5] = 255

and check that the padded versions behave as expected. What do you think?

[utkuozbulak]

Inspired from what you suggested as the test case Included the following test:

1- Create an empty tensor (image) with zeros

2- Change leftmost column to one.

3- Pad image with a constant value with a length of 1

4- Pad image with edge, symmetric or reflect (3 different cases, below only illustrated symmetric version)

Check the values of leftmost middle crop with the length of six (highlighted in the box above), if it is indeed symmetric. Then check the shape of the matrix if it adds up to what we are expecting (Here, I also test if the padding is a tuple instead of a single value in different versions of padding).

[fmassa]

This is awesome, thanks a lot for the contribution!