If we do this,

• direct users of rand_core might have issues (if they don't request the std feature and try to use missing functionality)
• indirect users of rand_core (i.e. users of rand etc.) shouldn't notice any differences, so long as the direct dependencies use the feature correctly
• this should help avoid problems like in Having rand 0.5 and rand 0.6 with default features turned off in the crate graph results in a compliation error #645, where the build happens to require std in rand_core (likely accidentally) but not in rand

There isn't a lot of functionality in rand_core which is enabled by the std feature, but there is a little:

• rand_core::Error::take_error is defined
• rand_core::Error implements std::error::Error
• std::io::Error implements From<rand_core::Error>
• RngCore implements std::io::Read

Of course, this is potentially a breaking change, so we should bump the version number to 0.4.0 and have a shim 0.3 → 0.4 which still enables std by default — except, since the change is unlikely to cause much breakage and we are already breaking some builds, I wonder if it is worth it?