Behavior of BFGS

Hi,

I'm not sure if this is the desired behavior, but BFGS directly stops if f(y0) == 0. As an example :

from optimistix import minimise, BFGS, rms_norm
import equinox as eqx
import jax.numpy as jnp


@eqx.filter_jit
def f(y, args):
    return jnp.sum(jnp.square(y))


N = 2


def f2(y, args):
    return f(y, args) - N


y0 = jnp.ones((N,))

res = minimise(
    f,
    BFGS(rtol=1e-13, atol=1e-13, norm=rms_norm),
    y0=y0,
    max_steps=1024,
)

print(res.stats["num_steps"])
assert jnp.allclose(res.value, jnp.zeros((N,)))

res = minimise(
    f2,
    BFGS(rtol=1e-13, atol=1e-13, norm=rms_norm),
    y0=y0,
    max_steps=1024,
)

print(res.stats["num_steps"])
assert jnp.allclose(res.value, jnp.zeros((N,)))

This is because then the termination condition is already validated at first step.
It's not really a big problem since you just need to offset the cost by a constant but it took me a while to figure out where the problem was coming from.
Maybe the initial value for f_info should be set to infinity, or maybe this should be mentioned in the doc ?

Reproduce the provided Python example first, then inspect the BFGS termination logic and initial f_info handling. Done means the behavior is resolved or documented, with both the f and f2 cases covered by a regression test.