Latest eccodes + workaround seg fault when saving

[lbdreyer]

Previously, we had a found a seg fault was being raised when trying to save out a large array.

This has now been mostly fixed in eccodes, but there is still an outstanding issue.
When trying to save out to GRIB, if the data is not float64, GRIB tries to cast it to float64. This is raising a seg fault!

As a temporary workaround, if we cast the data upfront, is seems to avoid the seg fault.

This PR is made up of 2 commits:

• the first commit unpins the eccodes requirements such that we get the latest eccodes, and also adds a test that explicitly tests saving int32, int64 and float32 (all of which would need casting to doubles). As you can see from the travis testng, this raises a seg fault
• the second commit adds the temporary workaround to cast the data to int64 upfront.

Note it turns out the seg fault was also being raised by the TestGDT5 but I thought I test specific to testing the data_section would be more helpful in future, rather than a rountripping test that just so happens to catch the issue.

This is has successfully been tested with the code that demonstrated the original seg fault from the two users who raised this. This has also been successfully tested with the iris-master tests (since there are still a couple grib-other fileformat interoperability tests.

[coveralls]

Coverage increased (+0.006%) to 92.549% when pulling 643487c on lbdreyer:eccodes_bug into 12b989b on SciTools:master.

[lbdreyer]

Thanks for merging @pp-mo !

[GoPavel]

Fortunately, seems Eccodes fixed segfault and will add that in the next release.
ecmwf/eccodes-python#30