`complex_rail_generator` generates invalid start position for trains on `env.reset()`

The first map generated by complex_rail_generator generates valid starting positions for trains, but resetting the env generates invalid starting positions for the trains, i.e. the trains are not on the tracks.

Code to reproduce (in jupyter)

import datetime

from PIL import Image
import matplotlib.pyplot as plt

%matplotlib inline


from flatland.envs.generators import (
    complex_rail_generator,
    random_rail_generator,
    empty_rail_generator,
)
from flatland.envs.rail_env import RailEnv
from flatland.utils.rendertools import RenderTool


number_of_agents = 2
width, height = 8, 8

env = RailEnv(
    width=width,
    height=height,
    rail_generator=complex_rail_generator(
        nr_start_goal=6, nr_extra=30, min_dist=10, max_dist=99999, seed=0
    ),
    number_of_agents=number_of_agents,
)
renderer = RenderTool(env)

# reset generates invalid starting positions
for _ in range(3):
    start_ts = datetime.datetime.now()
    env.reset()
    renderer.renderEnv(show=False)
    display(Image.fromarray(renderer.gl.getImage()))
    print(f"Generating map took {datetime.datetime.now() - start_ts}")

Tested with flatland 0.2.