Optimize dispatcher by using positional only arguments

[Ericgig]

Description
I optimized the dispatched by only supporting positional arguments for matrix to dispatch on:

The dispatcher was supporting any signature for dispatched functions.
Since each dispatchers is an instance of the Dispatcher class and cython does not support patching methods, it had to find the inputs matrix to convert from __call__(*args, **kwargs) and re-implemented the _bind method to do so.
But all our dispatched functions have the matrix as the first positional argument, by embracing this we get a nice speed up for small matrices.

Also when the specialization existed, it would not call it directly, but call a _constructed_specialisation that called ti. Adding another unneeded layer to each calls. This was changed so it would call the function directly.

For a 2x2 matrix:

	Before	After
data.add_dense	214ns	210ns
data.add (dense)	774ns	388ns
data.add[Dense, Dense]	541ns	182ns
data.add_csr	429ns	447ns
data.add (csr)	986ns	649ns
data.add[CSR. CSR]	744ns	401ns
numpy	473ns	-

Sadly it will not improve the benchmarks that much since the Qobj operations overhead is quite large:

	Before	After
Qboj + Qobj (dense)	2430ns	2070ns
Qboj + Qobj (csr)	3070ns	2330ns

[coveralls]

Coverage: 75.317% (-0.01%) from 75.328% when pulling d1160dc on Ericgig:enh_dispatcher into ab6a0d7 on qutip:master.

[hodgestar]

I left some small suggestions, but overall this is a great simplification.