I noticed when profiling my code that glm::fastLength's usage of glm::fastInverseSqrt pulls in the generic detail::compute_inversesqrt, which just calls out to sqrtf(). This was the case regardless of the precision of the vectors involved.

This change resolves that, by ensuring for usage with type float that the optimized detail::compute_inversesqrt specialization is used. With this change, no call is made to sqrtf() when calling glm::fastInverseSqrt.

I also noticed that the version under ifdef __CUDACC__ already seemed to use lowp so this also brings the two cases closer together in terms of implementation.