What

Multinomial::pmf and Multinomial::ln_pmf now compute the distribution entirely in log space, so valid large-n inputs return finite values instead of NaN/+inf.

• pmf delegates to self.ln_pmf(x).exp() (the length/sum guards are preserved: ln_pmf still panics on a length mismatch and returns NEG_INFINITY when the counts don't sum to n, and NEG_INFINITY.exp() == 0.0).
• ln_pmf builds the multinomial coefficient directly with factorial::ln_factorial (ln n! - Σ ln x_i!) instead of factorial::multinomial(...).ln().
• Zero-count categories are skipped in the probability sum so an allowed zero probability no longer produces 0 * ln(0) = NaN.

Why

factorial::multinomial returns the coefficient itself (it exponentiates the stable log value before returning), which overflows f64 to +inf once the log exceeds ~709.78. For n = 2000:

• pmf computed inf * prod(p_i^x_i), and since the probability product simultaneously underflows to 0.0, the result was inf * 0.0 = NaN.
• ln_pmf computed inf.ln() = +inf.

Both are wrong for perfectly valid inputs; SciPy returns finite values for the same parameters. Fixes #370.

Testing

Added test_pmf_large_n_no_overflow (n=2000, p=[0.5, 0.5], x=[1000, 1000]), which asserts the pmf is finite and matches the lgamma-computed reference values (pmf ≈ 0.0178390111, ln_pmf ≈ -4.0263675824). Verified it fails on the previous implementation (assert!(p.is_finite()) panics) and passes after the fix. The existing test_pmf cases still pass under the log-space implementation.

$ CARGO_BUILD_JOBS=2 cargo test -p statrs multinomial
test result: ok. 10 passed; 0 failed; 0 ignored; 0 measured; 716 filtered out

cargo clippy -p statrs --all-targets -- -D warnings is clean; cargo fmt --check reports no diffs in the touched file (the only fmt diffs on this tree are pre-existing in geometric.rs).