Improve rotation-to-operation flow

[Strilanc]

Feels really clunky. Easy to miss operations when not watching the screen. Hard to undo an operation sequence. Hard to hit Hadamard without hitting something in between. No correction of yaw drift.

Ideas:

• Only do "what's that operation?" inference when parked.
• Only do non-Pauli operations when parked.
• Don't ever infer an operation until an interaction with another qubit.
• Drift yaw towards rounded operation when parked.
• Beep?
• Handle axes separately.

[Strilanc]

I've decided to instead just not try to round the rotations into operations.