[aarch64] Fix fmin and fmax on aarch64

[shibatch]

vmaxnum_v*_v*_v* and vminnum_v*_v*_v* are helper functions for utilizing instructions for choosing larger and smaller elements from two given vectors, and they are introduced in the following PR.

#109

It is said that vmaxnmq and vminnmq on aarch64 are IEEE754-conformant, but I recently found that handling of signalling NaN by these instructions is not conforming to the specification of fmin and fmax functions in the ANSI C standard.

This patch fixes that problem.

[fpetrogalli]

Would it be possible to add a regression test? Also, do you mind linking some official information about this problem?
Moreover, we introduced the SLEEF_SINGLE_MINMAXNUM_AVAILABLEand SLEEF_DOUBLE_MINMAXNUM_AVAILABLE in #109. If we opt for the solution in this PR, I think we should remove those macros and all the changes related to that.

[shibatch]

Do you think it is better to test a signalling NaN input to all functions?

[fpetrogalli]

Do you think it is better to test a signalling NaN input to all functions?

No, I think it is enough to add a regression test for this particular PR.

[shibatch]

apt-get update at travis did not work anymore, so I added lines for updating expired keys and removing an apt source relating to mongo-db following the comments in travis-ci/travis-ci#8554 to .travis.yml.

[shibatch]

apt-get update problem seems to be fixed now.

travis-ci/travis-ci#9037

[shibatch]

As for the operation carried out by the aarch64 instructions, they are the minNum and maxNum operations defined in section 5.3 and 6.2.3 of IEEE Std 754™-2008.

http://ieeexplore.ieee.org/document/4610935/

The IEEE standard specifies that signaling NaNs should be propagated in these operations.

On the other hand, the C99 standard does not distinguish between a quiet NaN and a signaling NaN.

Since SLEEF is a substitute of the C99 math library, propagation of a signaling NaN is not desirable.