Dia data layer

[Ericgig]

Description
New sparse type: Dia.
Optimal for diagonal operators such as qeye, destroy. But it's quite bad for vectors ('bra', 'ket').

For a tri-diagonal matrix, it's about 25% faster than CSR for oper @ dense.
It can do an very efficient dense @ dia, while CSR is only efficient csr @ dense, not the opposite.

This include most specializations, change the default in operators, add tests.

The implementation mostly match the one in scipy.
The matrices are stored in two arrays, the diagonals offsets shape=num_offsets and the data shape=[num_offsets, ncols]. The diagonals are lined so that for each entry, the column in the data array match the matrix column.
The main differences with scipy's implementation are fixed dtype and the shape of the data array being fixed. (scipy accept longer diagonals and ignore entries outside the matrix.)

[coveralls]

coverage: 84.68% (+1.5%) from 83.199% when pulling 6e00623 on Ericgig:feature.diag into 4f29ae5 on qutip:master.

[Ericgig]

Here are some timing:
Operators are all 100x100

Sparse matrices composed of one diagonal:

	C	Fortran
dia @ dense	14.2 µs	14.7 µs
dense @ dia	14.2 µs	12.8 µs
csr @ dense	49.8 µs	33.3 µs
dense @ csr	165 µs	113 µs

Here for matrices with 5 diagonals:

	C	Fortran
dia @ dense	67 µs	63 µs
dense @ dia	60 µs	55 µs
csr @ dense	236 µs	85 µs
dense @ csr	270 µs	218 µs

[hodgestar]

I didn't do a thorough review, but we have lots of tests for the data layer and I can see you extended them to the new dia layer.

Happy for this to be merged.

Should we add it to the QuTiP benchmarking suite?

[Ericgig]

Yes, let's add it to the benchmarks.