Incorrect (or at least inconsistent) T cost for a TwoBitCSwap gate.

[wjhuggins]

To summarize a discussion I had with @tanujkhattar, @mpharrigan, and @NoureldinYosri:

We currently count the T and Clifford costs for a TwoBitCSwap (Fredkin) gate using the construction from https://arxiv.org/pdf/1308.4134. However, we can implement this gate using two CNOT gates and a Toffoli gate, which has a four T-gate construction using ancilla qubits (https://arxiv.org/pdf/1212.5069).

I think we should default to the ancilla-assisted construction for the Fredkin gate.

The simplest solution might be to no longer treat a Fredkin gate as a leaf node at all and instead decompose it into a Toffoli gate and two CNOT gates. If not, then we should at least be consistent in how we treat the resource counts for Fredkin and Toffoli gates.

[wjhuggins]

This is connected to #1121, which is related to the observation that we can do compute / uncompute pairs for Toffoli and Fredkin gates more cheaply than a naive implementation.

[mpharrigan]

As a first step, I'd recommend writing a simple circuit as a bespoke method like

https://github.com/quantumlib/Qualtran/blob/b3000cb3a4f10f99d260f487c21792cc1b9a0ab1/qualtran/bloqs/basic_gates/swap.py#L156

and updating the references section to highlight the disparity.

As a leaf bloq, we can change the T cost of a cswap on the fly. See the ts_per_cswap argument to GateCounts.total_t_count and GateCounts.to_legacy_t_complexity. Those default to 7. It can be changed to 4. Although maybe the "legacy" method should keep using 7 since it's been that way for a while