Use Extended/ExtensiblePoint more appropriately

[daxpedda]

Currently most mathematical operations on the twisted curve are implemented on ExtensiblePoint and take an ExtendedPoint as parameter. However, these operations often just convert to an ExtendedPoint internally (by multiplying T1 and T2).

This PR moves operations to "where they belong" to make sure we don't unnecessarily call these conversions. This achieves removing 1 multiplication from Decaf point add/sub operations from previously 9 in total, so a ~10% performance improvement.

There are a couple of other operations that are affected negligibly:

• Removes 7 multiplications from the lookup table construction for Edwards scalar multiplication. Reuse Edwards windowed scalar multiplication for Decaf #1303 switches Decaf scalar multiplication to the same algorithm.
• Removes 226 multiplications from Decaf scalar multiplication. While this is very major, Reuse Edwards windowed scalar multiplication for Decaf #1303 switches to a different unaffected algorithm.
• Removes 1 multiplication from Edwards point scalar multiplication when converting back to un-twisted form.