Fix testing strategies

[jclark-ce]

Address #360, changes are:

• Made the gates strategy filter gates ahead of time rather than relying on hypothesis
• Fixed frozen_gates to return only non-constant gates when requested, as well as accounting for None constant inputs
• Switched floats calls to use the default width argument, which drastically reduces the rejection rate for param lists (this defaults to 64-bit floats, which should produce valid params as far as I can tell)

[jclark-ce]

I guess I must've forgotten to re-run the tests before making the floats call change. Anyway, the errors in the last test were due to nan popping up in the results with large param values (which caused the tests to fail since nan is not equal to itself). The new version clamps the floats calls to the range $[-2^{16},2^{16}]$, which is roughly equivalent to using 16-bit floats (but without the failure rate).

Having actually run tests on the new version, it looks like they all pass with the exception of test_power_gate, which might be an actual bug, either in PowerGate or in the test_power_gate fixture.