Add statistics=false option to skip summary statistics in schema

[matthieugomez]

Computing a ContinuousTerm's mean/var/min/max takes several O(n) passes over each column, which can take as long as fitting the model itself (#222). Removing the fields is breaking (#223), so this adds a non-breaking opt-out instead, schema(f, data; statistics=false), which fills them with NaN placeholders.

Here are some performance checks on this branch:

using StatsModels
N = 10_000_000
df = (y = rand(N), x1 = rand(N), x2 = rand(N))
f = @formula(y ~ x1 + x2)

@time schema(f, df)                    # 0.126 s
@time schema(f, df, statistics=false)  # 0.7 µs
@time [df.x1 df.x2] \ df.y             # 0.112 s (the regression itself, for scale)