Check last point when using best so far minimiser

[ColCarroll]

Fixes #33

There is some unnecessary looking lines

        best_f, best_aux = fn(state.best_y, args)
        best_loss = self._to_loss(state.best_y, best_f)

where I would expect to just use state.best_loss, but the test doesn't pass without it!

[patrick-kidger]

Agreed, I definitely don't understand these lines. What do you think is going on?

[patrick-kidger]

Can we remove this line?

[patrick-kidger]

Thank you for the fix! Always happy to squash bugs :)

[ASKabalan]

@patrick-kidger Hello 👋

Can this be merged

This issue causing a bug for solvers that converge too quickly

import jax.numpy as jnp
import optimistix as optx

def fn(y , arg):
    return jnp.sum((y - 2.0)**2) + jnp.sum(y **2) * 0.0

DIM = 10
key = jax.random.PRNGKey(42)
y0 = jax.random.normal(key, (DIM,))

solver = optx.BestSoFarMinimiser(optx.BFGS(atol=1e-8 , rtol=1e-8))
sol = optx.minimise(fn, solver, y0, max_steps=1000)
print(f"Solution found: {sol.value}")
print(f"Best loss: {sol.state.best_loss}")
print(f"Loss using last params {fn(sol.value , None)}")

This PR solves it