ODE solvers: defer RT-safe derivative API until Mojo supports parametric traits

[mjsyts]

The current [working] RK2/RK4 implementation creates temporary InlineArray allocations inside step when calling fn_deriv, which is not ideal for real-time audio. The correct fix as I see it is a Differentiable[num_dims: Int] trait that structs implement, allowing the derivative function to live alongside its runtime state (e.g. frequency, damping). This was designed but blocked on Mojo not yet supporting parametric traits. When that feature lands, refactor RK2/RK4 to take a Differentiable conforming struct as a parameter instead of fn_deriv.

trait Differentiable[num_dims: Int]:
    fn derivatives(mut self, state: InlineArray[Float64, num_dims], mut derivs: InlineArray[Float64, num_dims]) -> None

struct RK2[num_dims: Int, T: Differentiable[num_dims]](Copyable, Movable):
    var state: InlineArray[Float64, Self.num_dims]
    var world: World
    var dt: Float64
    var k1: InlineArray[Float64, Self.num_dims]
    var k2: InlineArray[Float64, Self.num_dims]
    var temp_state: InlineArray[Float64, Self.num_dims]
    var differentiable: T

    fn __init__(out self, world: World, differentiable: T):
        self.world = world
        self.dt = 1.0 / world[].sample_rate
        self.state = InlineArray[Float64, Self.num_dims](fill=0.0)
        self.k1 = InlineArray[Float64, Self.num_dims](fill=0.0)
        self.k2 = InlineArray[Float64, Self.num_dims](fill=0.0)
        self.temp_state = InlineArray[Float64, Self.num_dims](fill=0.0)
        self.differentiable = differentiable

    fn step(mut self):
        self.differentiable.derivatives(self.state, self.k1)
        for i in range(Self.num_dims):
            self.temp_state[i] = self.state[i] + self.k1[i] * (self.dt / 2.0)

        self.differentiable.derivatives(self.temp_state, self.k2)
        for i in range(Self.num_dims):
            self.state[i] = self.state[i] + self.k2[i] * self.dt