Only normalize if initial state is normalized.

[Ericgig]

Description
We had a few issues (#2415, #2426) where an initial state that has not a trace of 1 was used and the new normalization of density matrices resulted in wrong results.

This change the normalization to happen only when the initial state is already normalized, expecting that when a non-normalized initial density matrix in used in mesolve, etc., it is a conscious choice and the solver should not fight it.

Related issues or PRs
fix #2426

[coveralls]

coverage: 86.266% (+0.006%) from 86.26%
when pulling 528200b on Ericgig:misc.solver.state0_norm
into ad05a22 on qutip:master.

[BoxiLi]

I actually have a question about this automatic normalization.

Do we really need this to hide the numerical errors? Wouldn't it be better to just expose the error since it will be proportional to atol anyway (and if not, something went wrong)? Which group of users will this auto-normalization benefit?

Sometimes an unnormalised state at the ends can also indicate that there is something wrong with the Hamiltonian, e.g., it is not hermitian due to some typos.

[Ericgig]

We mostly added it to mesolve since we always had it for sesolve.
Numerical error is not well understood by all users and we can get issues that the solver are wrong because the trace, norm, populations etc. is not constant.