Deregistering in RPC handler

[mdorier]

If an RPC handler calls HG_Deregister to deregister itself, a subsequent HG_Respond will cause a segfault, with the following backtrace:

#0  hg_set_struct (hg_handle=hg_handle@entry=0x55fe46cef810, hg_proc_info=0x594925d9ff758133, op=op@entry=HG_OUTPUT,
    struct_ptr=struct_ptr@entry=0x7f94fce16ef0, payload_size=payload_size@entry=0x7f94fce16ec0, more_data=more_data@entry=0x7f94fce16ebf "")
    at /tmp/mdorier/spack-stage/spack-stage-mercury-2.2.0-ftf2olh6pdj2dt2vn4fhcxifrucwuuba/spack-src/src/mercury.c:577
#1  0x00007f94fe9e2287 in HG_Respond (handle=handle@entry=0x55fe46cef810, callback=callback@entry=0x7f94fea3e600 <margo_cb>,
    arg=arg@entry=0x7f94fce16f50, out_struct=out_struct@entry=0x7f94fce16ef0)
    at /tmp/mdorier/spack-stage/spack-stage-mercury-2.2.0-ftf2olh6pdj2dt2vn4fhcxifrucwuuba/spack-src/src/mercury.c:2044

Ideally, an error in HG_Respond should be returned.

I haven't checked if other functions (like HG_Get_input/output) succeed, fail, or crash, but maybe it would be good to ensure a well-defined behavior of functions that use a handle associated with an RPC that has been deregistered.

[soumagne]

Handles cache RPC info on server to prevent extra lookups from RPC ID hash table, we'll need to either refcount entries in hash table or ensure that we cannot deregister an RPC until all RPCs under that ID have been executed.

[soumagne]

Linked to #534