fix: compute FisherSnedecor pdf in log space for large freedoms

[teddytennant]

FisherSnedecor::pdf evaluates ((d1 x)^d1 * d2^d2) / (d1 x + d2)^(d1+d2) with powf. At ordinary in-range parameters those powers overflow independently, so pdf comes back 0 or NaN and ln_pdf (which is just pdf(x).ln()) cannot recover.

FisherSnedecor::new(80.0, 80.0).unwrap().pdf(1.0) // was 0.0
FisherSnedecor::new(100.0, 100.0).unwrap().pdf(1.0) // was NaN

F(80, 80) at the mode x = 1 is a textbook density of about 1.78, not a tail.

Compute ln_pdf from the log of the documented formula (ln_beta already exists) and take pdf = ln_pdf.exp(). Same class of fix as the Hypergeometric / Multinomial pmfs.

A few existing small-dof fixtures differ by about 1 ULP from the old product-then-sqrt path, so those tolerances moved one digit (for example 1e-16 to 1e-15). That is not how the overflow cases are tested; those have their own expected values.

cargo test --lib is green (792 passed).

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 95.07%. Comparing base (92819b6) to head (0a84d2b).

Additional details and impacted files

@@ Coverage Diff @@## main #448 +/- ##
=======================================
Coverage 95.07% 95.07% =======================================
Files 62 62 Lines 14191 14203 +12 =======================================
+ Hits 13492 13504 +12 
Misses 699 699 

☔ View full report in Codecov by Harness.
📢 Have feedback on the report? Share it here.

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.