Fix errors in hyperbolic function crossover

[MattStephanson]

Fixes #1059.

• Need to introduce a floating point to avoid integer truncation.
• The exp(-x) term needs to be smaller than ULP/2, which requires x > (p+1)*ln(2)/2. But NBITS is the number of physical, not effective, bits, so (NBITS+2) is needed. Could also use numeric_limits<T>::digits+1.

[ghost]

All CLA requirements met.

[StephanTLavavej]

Thanks! We're a bit busy at the moment with VS 2019 16.8 Preview 3 but we'll try to review your PRs soon. 😸

[cbezault]

I'm convinced that this is correct but could you add a (very) small test for this change? It could be as simple as using the examples that @statementreply put in #1059 as a sanity check.

[MattStephanson]

I'm convinced that this is correct but could you add a (very) small test for this change? It could be as simple as using the examples that @statementreply put in #1059 as a sanity check.

Added test case. Sorry if my rebase messed things up, I can reset to fc95b36 if need be.

[StephanTLavavej]

Thanks, this looks good to me! I'll push a couple of minor test changes.

[StephanTLavavej]

Thanks for fixing these long-standing correctness bugs, and congratulations on your first microsoft/STL commit! 😺 🚀