Use TransmuteDims

[mcabbott]

This replaces orient, rview (which reshaped), dropdims, and permutedims with transmute from https://github.com/mcabbott/TransmuteDims.jl

Known problems:

• @strided may not work with this yet. I think this is fixed.
• tracking of lazy/not is broken, |= & nolazy Maybe fixed, too.

[mcabbott]

The examples of #28 will become:

julia> @pretty @reduce C[a,b,d] := sum(c) A[a,b,c] * B[a,c,d] # Fast
begin
    local ram = transmute(B, (1, nothing, 2, 3))               # reshape
    C = transmute(sum(@__dot__(A * ram), dims = 3), (1, 2, 4)) # reshape
end

julia> @pretty @reduce C[b,a,d] := sum(c) A[a,b,c] * B[a,c,d] # Not fast
begin
    local ram = transmute(A, (2, 1, nothing, 3))    # one wrapper
    local tapir = transmute(B, (nothing, 1, 3, 2))  # one wrapper
    C = transmute(sum(@__dot__(ram * tapir), dims = 4), (1, 2, 3)) # reshape
end

and that of #25:

julia> @pretty @reduce D[m,a] := sum(p) C[p,a] + L[p,m]
begin
    local octopus = transmute(C, (nothing, 2, 1))
    local deer = transmute(L, (2, nothing, 1))
    D = transmute(sum(@__dot__(octopus + deer), dims = 3), (1, 2))
end

Both of these benefit from avoiding needing wrappers of wrappers, as their reshape(PermutedDimsArray(... can be one lazy transmute.