Performance improvements for trig functions with @fastmath

[sostock]

#745 improved the accuracy of cis for quantities in ° (degrees), but this unfortunately comes with a performance hit.

Since the @fastmath implementation for cis (and other trig functions) falls back to the normal implementation for some types, this also impacted the performance of @fastmath cis(…). Since users of @fastmath most likely care more about performance and are willing to sacrifice a little numerical accuracy, this PR improves @fastmath performance for sin, cos, tan, sincos, and cis.

Performance with @fastmath is now at least as good as before #745 (in some cases better), except when using BigFloat with °: Before this PR, the conversion from degrees happened with uconvert, which for ° → rad only has Float64 precision. It now uses deg2rad for that, which is a bit slower but has BigFloat precision. I think that this is fine, since you probably want BigFloat precision if you are using BigFloat.