Clarify error handling

[krlmlr]

#1051 (comment)

[maelle]

@schochastics I wonder whether this belongs to the ISC milestone, is it related to clarifying what's an R vs C responsability?

[szhorvat]

This is about ensuring that the glue code in the R interface plays nice with the error handling of the C core. In particular, longjumps MUST NOT happen except at the top level. Rf_error does a longjump by default.

Basically this boils down to writing down some rules about when we use IGRAPH_CHECK vs IGRAPH_R_CHECK, and ensuring that we're following these rules. In other words, decide what function is considered top level and what isn't, and make sure that top-level functions are never called from non-top-level ones. Top-level uses IGRAPH_R_CHECK, everything else uses IGRAPH_CHECK.

Related: #514

Currently, the error handling is a bit of a mess, and when certain errors happen, the R interface WILL blow up in some way, or have its internal state corrupted. The reason why we don't see this is that most such errors are unlikely (such as out-of-memory).