Discrete Baker's Map

Pyth, 25 19 18 bytes

msC_dcs_Cmck/lk2Q2

Online demonstration. It uses an 2D-array of chars.

Array of strings is one char longer (19 bytes). Online demonstration

Explanation:

         m      Q    map each string k in input:
            /lk2        calculate half the line-length: len(k)/2
          ck/lk2        chop k into pieces of length len(k)/2
                        results in two pieces
        C            zip the resulting list
                     results in a tuple ([first half of strings], [second half of strings])
       _             invert the order ([second half of strings], [first half of strings])
      s              sum (combine the two lists to a big one
     c           2   chop them into tuples
m                          for each tuple of strings: 
 sC_d                        invert, zip, and sum

The last part is a bit confusing at first. Let's assume we have the tuple ['DEF', 'JKL'] (I use the example from the OP).

    d  (('D', 'E', 'F'), ('J', 'K', 'L'))   just the pair of strings
   _d  (('J', 'K', 'L'), ('D', 'E', 'F'))   invert the order
  C_d  [('J', 'D'), ('K', 'E'), ('L', 'F')] zipped
 sC_d  ('J', 'D', 'K', 'E', 'L', 'F')       sum (combine tuples)

Julia, 136 bytes

A very straightforward implementation. Not a particularly competitive entry, but it was fun!

A->(s=size(A);w=s[2];u=w÷2;C=vcat(A[:,u+1:w],A[:,1:u]);D=cell(s);j=1;for i=1:2:size(C,1) D[j,:]=vec(flipdim(C[i:i+1,:],1));j+=1end;D)

This creates a lambda function that accepts a 2-dimensional array as input and returns a transformed 2-dimensional array.

Ungolfed + explanation:

function f(A)

    # Determine bounds
    s = size(A)          # Store the array dimensions
    w = s[2]             # Get the number of columns
    u = w ÷ 2            # Integer division, equivalent to div(w, 2)

    # Stack the right half of A atop the left
    C = vcat(A[:, u+1:w], A[:, 1:u])

    # Initialize the output array with the appropriate dimensions
    D = cell(s)

    # Initialize a row counter for D
    j = 1

    # Loop through all pairs of rows in C
    for i = 1:2:size(C, 1)

        # Flip the rows so that each column is a flipped pair
        # Collapse columns into a vector and store in D
        D[j, :] = vec(flipdim(C[i:i+1, :], 1))

        j += 1
    end

    return D
end

To call it, give the function a name, e.g. f=A->(...).

Example output:

julia> A = ["A" "B" "C" "D" "E" "F";
            "G" "H" "I" "J" "K" "L";
            "M" "N" "O" "P" "Q" "R";
            "S" "T" "U" "V" "W" "X"]
julia> f(A)

4x6 Array{Any,2}:
 "J"  "D"  "K"  "E"  "L"  "F"
 "V"  "P"  "W"  "Q"  "X"  "R"
 "G"  "A"  "H"  "B"  "I"  "C"
 "S"  "M"  "T"  "N"  "U"  "O"

julia> B = ["H" "e" "l" "l";
            "!" " " " " "o";
            "d" " " " " "-";
            "l" "r" "o" "W"]
julia> f(B)

4x4 Array{Any,2}:
 " "  "l"  "o"  "l"
 "o"  " "  "W"  "-"
 "!"  "H"  " "  "e"
 "l"  "d"  "r"  " "

And proof that it can be arbitrarily chained:

julia> f(f(B))

4x4 Array{Any,2}:
 "W"  "o"  "-"  "l"
 "r"  " "  " "  "e"
 "o"  " "  " "  "l"
 "l"  "!"  "d"  "H"

Suggestions are welcome as always, and I'll happy provide any further explanation.

CJam, 25 24 bytes

qN/_0=,2/f/z~\+2/Wf%:zN*

Straight forward spec implementation. Explanation:

qN/                       "Split input by rows";
   _0=,2/                 "Get half of length of each row";
         f/               "Divide each row into two parts";
           z              "Convert array of row parts to array of half columns parts";
            ~\+           "Put the second half of columns before the first half and join";
               2/         "Group adjacent rows";
                 Wf%      "Flip the row pairs to help in CW rotation";
                    :z    "Rotate pairwise column elements CW";
                      N*  "Join by new line";

Try it online here