According to the read manpage a read could return less bytes then
requested due to an interrupt or an error.

We now try to repeat the read operation until all the existent data have
been read correctly.

It is possible that we read less bytes then requested cause the actual
file doesn't have so much.

The ChunkStorage class now store the chunksize in order to perform
runtime checks upon chunks bundaries.

Mercury now support shared memory autorouting.

A single margo instance can be initialized and it will handle both
shared memory communication and remote ones.

If the endpoint of the RPC is local mercury will automatically use
shared memory.

Since there is only one margo instance all the duplicated code for
rpc/ipc have been unified and simplified considerably.

------

The way in which client contact the server has changed.

 - Server initializes its own margo instance and generate the endpoint
communication string using `HG_Addr_self`.
 - This endpoint description string is written on the pid file
 - When the client library loads it will fetch the server endpoint
description from the pid file and will use that to contact the server.

Searching for mercury dependencies is useless. In fact if a library has not been linked and enabled while mercury was built, we won't be able to use it.

Moreover, mercury ships with a make module that provides which links flags needs to be used. Thus there is no need to search for mercury dependencies.

Also removed mercury_cci_verbs_lookup patch, because it has been merged upstream [1]

[1]: https://github.com/mercury-hpc/mercury/pull/206

on centos the cmake 3 binary is called `cmake3`. In the compile script
we try to use the most recent version of cmake that is available on the
system.

This will allow to prevent this mercury installation issue:
https://github.com/mercury-hpc/mercury/issues/224

I couldn't figure out why this is happening, but the inline function
cannot be found at runtime

ensure rpc_addresses object have been initialized before we use it for
the first time

margo, argobots and mercury have been updated. The abt-snoozer dep
required by margo it is not needed anymore.

RocksDB has been also updated to the lastest release

Instead of using custom variable to search for installed libraries
we can rely on the standard `CMAKE_PREFIX_PATH` [1]

[1]: https://cmake.org/cmake/help/v3.8/variable/CMAKE_PREFIX_PATH.html#variable:CMAKE_PREFIX_PATH

compile/download scripts need to exit with positive error code on wrong parameters

loggers can be now configured at execution time, by using environment
variables for both server and client.

Server:
 - ADAFS_LOG_LEVEL
 - ADAFS_DAEMON_LOG_PATH

Client:
 - ADAFS_LOG_LEVEL
 - ADAFS_PRELOAD_LOG_PATH

This allows us to quickly see if the file system can provide basic functionality without breaking.

Since we introduced the `initialized` flag in the preload context, it is
not needed anymore to check for ld_is_aux_loaded flag.
In fact ld_is_aux_loaded is true if and only if the `initialized` flag
is true.

The library intercepts IO functions from others library constructors
before our constructor have been executed.

All the interceptions triggered before that our constructor have been
executed are now forwarded to the glibc.

By moving the logger initialization outside of the passthrough
initialization we will be able to implements fopen interceptions.