fp: dedicated assembly for square

[mmcloughlin]

As a stopgap #32 implements square with multiply. We should have dedicated assembly for this that saves on 64-bit multiplies.

[mmcloughlin]

The observation is that we don't need to compute as many partial products due to symmetry. That is, we can write

x^2 = 2 * SUM_{i < j} x_i * x_j * 2^{b(i+j)}
    + SUM_i x_i^2 * 2^{2bi}

[mmcloughlin]

Status of work-in-progress implementation:

       ec3      go
sqr     66      60
mod     90      59

Montgomery reduction is far too expensive. Probably for two reasons:

• Not exploiting the structure of the prime p #95
• Excessive additions