Fix pdf/cdf domain handling in dist (StudentT negative cdf, Binomial overflow/infinite loop, off-support inputs)

[gaoflow]

While cross-checking statistics::dist against scipy 1.17.1 (all nine distributions, pdf + cdf, on and beyond their supports), I found the implementations agree to ~1e-7 on the happy path but four families break as soon as the input leaves it:

StudentT cdf is wrong on the entire negative half. The negative branch used self.cdf(-x) - 0.5 where the reflection is 1 - self.cdf(-x):

let t = StudentT(3f64);
t.cdf(-5.0);   // 0.4923  (scipy: 0.00770)
t.cdf(-1.0);   // 0.3045  (scipy: 0.19550)

Binomial pdf hangs the process for x > n and silently returns garbage from n ≈ 30. The pmf went through ops::C(n, m) in usize arithmetic. P(30, 15) already overflows, so in release C(30, 15) came out 14052247 instead of 155117520 — Binomial(30, 0.5).pdf(15.0) returned 0.0131 instead of 0.1445 with no error (debug builds panic instead). Worse, for m > n the reduction C(n, n - r) underflows usize and the recursion cycles: Binomial(10, 0.3).pdf(12.0) spins at 100% CPU forever in a release build. The pmf is now computed in log-space via the existing ln_gamma (exact-enough for any n, e.g. n = 100) behind a support check.

Binomial cdf panics at the top of its own support. cdf passed out-of-domain arguments to inc_beta: Binomial(10, 0.3).cdf(10.0) panicked with "Bad a or b in routine betai", same for k > n and k < 0, and fractional k was interpolated instead of floored. Now clamped (k < 0 → 0, k ≥ n → 1) and floored, matching scipy/R semantics.

Bernoulli pdf returned 1 − p for any x outside {0, 1} — pdf(-1.0), pdf(0.5), pdf(2.0) all gave 0.7 for p = 0.3. Now 0 off support.

At the root, ops::C/P now return 0 for r > n (so the r > n/2 reduction can't underflow and recurse forever) and use checked multiplication, so genuine overflow is a clear panic instead of silent wraparound in release. C uses the multiplicative formula, exact for every C(n, r) with n ≤ 62 on 64-bit; the old path broke around n = 30. factorial/double_factorial get the same checked treatment.

Tests: regression tests for every case above with scipy-cross-checked values, including the previously-hanging call, plus a new tests/ops.rs for C/P. Re-running the full sweep (9 distributions × pdf/cdf, 101 points) against this branch gives zero divergences from scipy.

cargo test (default and --no-default-features), cargo fmt, cargo clippy --all-targets are clean — the remaining clippy warnings are the pre-existing doctest fn main ones.

One behavior note to flag: C/P/factorial panicking on true overflow (instead of returning wrapped values in release) is a deliberate change; if you'd rather have a different convention there (saturating, Option), happy to adjust.