Disable bottleneck by default

[takluyver]

Description

Certain algorithms in bottleneck are less numerically stable than their numpy equivalents, causing e.g. arr.mean() to give surprisingly inaccurate answers if bottleneck is installed. I was just bitten by this, and it's not obvious what the source of the issue is. Everyone commenting on #7344 seems to agree that the more accurate option should be the default, even if it's slower.

Checklist

• Closes Disable bottleneck by default? #7344
• User visible changes (including notable bug fixes) are documented in whats-new.rst

[takluyver]

There are several tests failing because the rank method requires bottleneck. IMO, it makes sense for the use_bottleneck option to control what's used when another option is available (i.e. fall back to plain numpy), but I don't think it needs to disable methods where bottleneck is the only option (unless e.g. just importing the module causes issues, which is not the case).

[keewis]

for rolling tests I think you can explicitly enable bottleneck (the test already requires bottleneck)

[takluyver]

Thanks @keewis! The tests are passing now, the remaining CI failure looks like a random glitch.

[keewis]

looks good to me.

Let me think about the testing strategy, but unless there's any objections I'll merge tomorrow

[keewis]

thanks again, @takluyver. I didn't end up deciding anything regarding the test strategy, so I postponed that further.

[takluyver]

Thank-you!