Add an improved stratifier

[perlinm]

Add a new and improved stratifying method, and (at least tentatively) an option as to which stratification method to use.

The basic idea is to:

• iterate over operations in topological order,
• create new moments on an as-needed basis, and
• advance moments up/forward to absorb a new operation, whenever possible.

Improvements over the old stratifier (from #6013) include:

• Shorter circuits! New circuit depth is upper bounded by the old circuit depth.
• Linear runtime complexity with the number of categories (the old classifier is factorial, O((len(categories)+1)!)).
  • Corollary: ~5x reduced runtime on the example below, with 2 categories.

Test script (to run with a local version of cirq that contains this PR):

#!/usr/bin/env python3
import cirq
import time

circuit = cirq.testing.random_circuit(qubits=4, n_moments=100, op_density=1)
categories = [lambda op: len(op.qubits) == 1, cirq.CNOT]  # single-qubit ops, and CNOTs

for method in ["static", "dynamic"]:
    start = time.time()
    new_circuit = cirq.stratified_circuit(circuit, method=method, categories=categories)
    print("time:", time.time() - start)
    print("length:", len(new_circuit))
    assert cirq.approx_eq(circuit.final_state_vector(), new_circuit.final_state_vector())

NOTE: this PR is NOT COMPLETE at the moment. However, the PR is relatively long, so I thought to open it up for initial review. Apologies if that's not appropriate, in which case please let me know and I'll mark this PR as a draft for now.

The current TODO list is:

• Properly deal with measurement/control keys (assumed to be absent at the moment).
• Decide whether to remove the old stratifier. Since the new stratifier seems to be strictly better, I cautiously lean "yes".
• Add/modify unit tests (which may depend on whether the old stratifier is removed).

Tagging @tanujkhattar, who reviewed my last statifier PR (#6013).

Feedback greatly appreciated 🙂, as well as suggestions for how to cleanly deal with measurement/control keys.