Silence warnings from nanmin and nanmax.

[mdickinson]

When all of raw_value consists of NaNs (which seems like a legitimate use-case, given that we already allow some of the values to be NaNs), the nanmin and nanmax calls in ImageData.get_bounds emit a RuntimeWarning.

This PR avoids that RuntimeWarning by replacing nanmin and nanmax with fmin.reduce and fmax.reduce calls.

[coveralls]

Coverage remained the same when pulling 2c52195 on silence-nan-warnings into 3e52ecb on master.

[corranwebster]

Is it too much to ask for a test?

[corranwebster]

Also, can you update CHANGES.txt

[corranwebster]

Otherwise LGTM.

[mdickinson]

Is it too much to ask for a test?

We should definitely add a test. Please can I put that on my to-do list for later today?

[mdickinson]

Merging, with @corranwebster's permission.

[coveralls]

Coverage remained the same when pulling be1a821 on silence-nan-warnings into 3e52ecb on master.