Remove deleter for the kiva_backend property (which is readonly)

[mdickinson]

PR #1670 added deleters for various ETSConfig attributes. However, the kiva_backend property is readonly, with value derived from the toolkit attribute, so we shouldn't be providing a deleter for it. The right way to mock kiva_backend is to set the toolkit attribute (and in fact there's already a test for that).

This PR removes the deleter for the kiva_backend attribute.

[corranwebster]

I think that the del has use here: like the other del methods, it effectively resets the value of kiva_backend. Without this, a reset requires directly settting _kiva_backend to `None.

Unless there is a problem I am not seeing.

[mdickinson]

kiva_backend is just a derived property of toolkit, so you can do a reset by deleting toolkit (or setting it to another value).

[mdickinson]

Unless there is a problem I am not seeing.

Ah no, there's a problem I'm not seeing, which is that resetting toolkit should set _kiva_backend to None. Cache invalidation and all that ...

[mdickinson]

Okay, I think this needs a closer look; I'm not at all convinced that the current code (even before the deleters were added) is behaving as it should be. I'll add some new tests and rework.

What I think we want is that the value of kiva_backend is always consistent with (and can be deduced from) the value of toolkit. The easiest way to achieve that would be simply not to cache the value: the computation of the value isn't at all time-consuming (no imports involved, for example) and this isn't something we're going to be accessing often enough for the caching to be worthwhile.

[mdickinson]

Updated to make kiva_backend uncached.

[corranwebster]

I think that this is reasonable.