This makes it easier for a patch to dune-grid to support context
objects optionally.

Otherwise the tags assigned to elements and those to DDD will not
match: the (global) tags assigned to elements will increment for each
new MG. After a new MG is created, the mapping to already existing
tags used by DDD will break as the MG-side will use new tags.

The local rank is only known from the `PPIFContext` which is not
available here.

This allows removing the global object later.

`GetMemoryForObject` ensures the memory is zeroed and the code
currently expects that.  So we explicitly zero out the memory as well.

UG's heap only wraps malloc/free anyway.

GCC 5.4 does not like using `XferMode::${something}` as there is also
the `XferMode` function visible at the same time.

This commit moves the state from `XFER_GLOBALS` to the context
object.  Some global variables still remain, for example the lists
managed by sll.h and `theXIAddData` from cmds.cc.

The casts are necessary as the structs cannot be defined in the `xfer`
module as long as it is still dim-dependent (same as for the previous
change to the `join` module).

Sorting the forward list of messages means a new first item is chosen,
however the loop to free all XFERMSG objects in DDD_XferEnd will start
from the old list head and no longer frees all entries.  This was not
noticed before the previous commit as the free in DDD_XferEnd was a
no-op and memory would be released by the heap structure instead.

Alternatively one could return the new list head (or pass in a
modifiable reference), but this seems more complex then just skipping
the sorting step.  I doubt handling low resource environments in this
way is still necessary.

The casts are necessary as the structs cannot be defined in the `join`
module as long as it is still dim-dependent.

Also convert `JoinMode` to a scoped enumeration type.