pdf loses the tail in several distributions that exponentiate before dividing

[teddytennant]

While fixing #452 I checked whether other distributions share the shape that caused it. Several do.

The pattern is computing pdf as exp(...) / normalizer in value space. The numerator underflows or overflows before the density itself does, and dividing afterwards cannot recover it. Measured on main at 128c9ae:

Normal(0,1e-300).pdf(5e-299)        = 0        true = 5.40515e-244
Laplace(0,1e-300).pdf(1e-297)       = 0        true = 2.53798e-135
Exp(1e300).pdf(1e-297)              = 0        true = 5.07596e-135
Chi(160).pdf(45.0)                  = 0        true = 2.5387e-318
Levy(0,2e-297).pdf(1e-300)          = NaN
Gamma(150,1e10).pdf(1e-8)           = NaN
InverseGamma(150,1e-10).pdf(1e-12)  = NaN
Pareto(1e-200,5).pdf(1e-100)        = NaN
Gumbel(0,1).pdf(-1000.0)            = NaN

The three true values above are exact arithmetic at 120 digits. A density is never NaN, so those five rows are wrong whatever the reference value is.

These are not all one fix:

• Laplace and InverseGamma define ln_pdf as self.pdf(x).ln(), so there is no correct log-space form to delegate to. Both need one written, the same way fix: compute FisherSnedecor pdf in log space for large freedoms #448 did for FisherSnedecor.
• Chi and Gamma already branch to ln_pdf(x).exp() above a shape threshold (160 and 160.0). The threshold guards the shape parameter, but the overflow here comes from rate.powf(shape) and x.powf(freedom - 1.0), which blow up independently of it. Those thresholds want re-tuning rather than a new branch.
• Gumbel fails through a double exponential overflowing to NaN, which is a cousin of this rather than the same thing.
• Normal is probably the one worth doing first. It is the most used distribution here, its pdf_unchecked is the same expression as the LogNormal one, its ln_pdf_unchecked is already correct, and StudentsT::pdf delegates to it for freedom >= 1e8.

I did not roll any of these into #454. That would turn a one-file fix into a rewrite of a third of the distribution module. Happy to send them as separate PRs if you want them, and happy to be told some are not worth changing.