Consistent rounding in adjacency matrix functions

[szhorvat]

Rounding should work consistently and identically in adjacency matrix functions.

There is currently a difference in rounding behaviour between igraph_adjacency() and igraph_biadjacency(). Only the latter uses ceil().

[szhorvat]

Decisions to be made:

1. Do we guarantee any sort of behaviour at all when some entries are not integers? It is tempting not to, as it simplifies things and makes efficient implementations easier.
2. Do we issue a warning or error if there's a non-integer matrix entry?
3. Do we round using ceil(), floor(), round(), trunc() or plain cast (which I believe is equivalent to trunc())? Currently, igraph_adjacency() does a plain cast (not documented) and igraph_biadjacency() uses ceil() (documented). Note that ceil(), floor() and round() are slow while trunc()/cast are fast (by a large margin).
4. When operating together lower/upper triangular elements, do we round before or after the operation? Currently it's done before.

[szhorvat]

My suggestions:

1. Guarantee a behaviour, but prioritize integer values. I.e. don't try to implement a "nice" behaviour for non-integer.
2. No strong opinion, leaning towards a warning (or not checking at all).
3. Use a plain cast, which is equivalent to trunc(). GCC/Clang generate the same code for trunc() and a cast. MSVC still generates a call to trunc(), which is likely much slower. This will be a breaking change for igraph_biadjacency(), but it brings consistency with igraph_adjacency().
4. There is no explicit addition here, round before.

[ntamas]

1. Agreed
2. A warning would be nice.
3. I don't think the performance implictions are significant here, except for maybe sparse matrices. Plain cast is okay for me.
4. Whichever is easiest to do in code, there's no real significance. If we round after, then it's only one operation instead of two, but it might complicate the code and there's no reason for micro-optimization here.