Avoid unnecessary heap allocation of SLICE objects

[wlevine]

This partly fixes #541. Here are the results of Lokesh's benchmark with the patch:

       user     system      total        real
Matrix getter  1.960000   0.000000   1.960000 (  1.961916)
NMatrix getter  1.570000   0.000000   1.570000 (  1.583510)
Matrix setter  1.680000   0.000000   1.680000 (  1.688569)
NMatrix setter  3.070000   0.000000   3.070000 (  3.077533)

#[] is probably fast enough now. #[]= is a lot faster than before, but may still need more work.

@mohawkjohn I think this is the right way to do this, but I'd appreciate it if you take a look.

[v0dro]

Why are the tests failing?

[translunar]

@v0dro Looks like the failures relate to #522 and #538, or at least resemble those. I can't see how any of @wlevine's changes would have caused this, though — can you?

[v0dro]

@mohawkjohn I think you're right. Those other bugs will need fixing for the build to pass.

[translunar]

@wlevine Can you please rebase against master? I think we've fixed the Travis problems.

[wlevine]

@mohawkjohn Done!