performance: Provide a fast subtle.ConstantTimeCopy alternative for more platforms

[Yawning]

As the scalar basepoint multiply implementation is based around using a table that stores pre-computed results, a fast constant time copy is required for optimal performance.

The amd64 target has an optimized implementation written in assembly language.
All other platforms will use unsafe (falling back to subtle.ConstantTimeCopy) once #19 is merged, but could be improved further.

• Merge internal/ge25519: Use the unsafe moveConditionalBytes on more targets #19.
• Figure out which architectures are sufficiently indifferent to alignment such that the runtime library fallback never needs to be used.
• Add assembly implementations for commonly used targets.