Categorical has poorly defiend behavior when some events are 0 probability

[YeungOnion]

Was looking to have more function implement try_inverse_cdf instead of wrapping the panicking version with Ok, found this.

Categorical::locate (used by both inverse_cdf and try_inverse_cdf) does:

    self.norm_cdf.binary_search_by(|v| v.partial_cmp(&x))

But this won't match the standard quantile definition inf{x | F(x) >= p} if this cdf isn't monotone because there's not an ordering if some categories have 0 weight - a weird but not outright unreasonable case.

Reproducer:

let norm_cdf = vec![1.0, 1.0, 1.0]; // categories [1.0, 0.0, 0.0]
norm_cdf.binary_search_by(|v| v.partial_cmp(&1.0).unwrap()); // Ok(1), want 0

let norm_cdf = vec![0.0, 0.0, 1.0]; // categories [0.0, 0.0, 1.0]
norm_cdf.binary_search_by(|v| v.partial_cmp(&0.0).unwrap()); // Ok(1), want 0,

let norm_cdf = vec![0.25, 0.5, 0.5, 0.5, 1.0]; // interior tie
norm_cdf.binary_search_by(|v| v.partial_cmp(&0.5).unwrap()); // Ok(2), want 1

inverse_cdf/try_inverse_cdf currently reject x == 0.0 and x == 1.0 outright, which happens to route around the two boundary cases above, but does nothing for interior ties (the third example).