Projected Gradient Descent Solver Interface

[cohnt]

Projected gradient descent (PGD) is a strategy for solving constrained optimization problems -- one alternates between taking a step down the gradient of the objective function, and projecting onto the feasible set. In some contexts, this optimization strategy is preferable to using a second order solver like SNOPT or IPOPT. Having a native PGD solver in Drake would particularly be useful if we want to use it for the rounding stage of GcsTrajectoryOptimization.

I'm assigning to myself for now, but probably won't be able to get to this for some time. If someone wants it sooner, I'm happy to hand it off. Minimum features to include:

• User can specify solver used for the projection step
• User can manually specify a gradient function, instead of computing with autodiff/symbolic
• Backtracking line search

Additional features that might be nice:

• Specifying the affine hull in closed-form (can simplify the projection, or render it moot altogether)
• Momentum, and various other tricks that can lead to faster convergence

cc @shrutigarg914 @RussTedrake