Alternative rotation compilations/costs

[fdmalone]

Currently our T costs derive from https://arxiv.org/pdf/1404.5320.pdf, which gives a mean cost of 1.149 log2
(1/ε) + 9.2 for a target precision of ε.

Many papers use a phase gradient gate instead for rotations (appendix A of https://doi.org/10.1103/PRXQuantum.1.020312) which requires b_r - 2 Toffolis for b_r bits of precision (it uses the first method to prepare a phase gradient register first and then adds into this). This is often significantly cheaper.

Need to think about rotations.

[mpharrigan]

cc @NoureldinYosri who has been thinking about this

[NoureldinYosri]

we currenly have a unified approach to the vanilla rotation models RotationLogarithmicModel and for constant cost models ConstantWithOverheadRotationCost both of which implement the same interface of RotationCostModel

Do we need to do more for this?

[mpharrigan]

Related: #1250

[mpharrigan]

Going to close this in favor of #1250 or open new issues to track specific asks