fix returned k when both tails give NA

[avehtari]

When using pareto_smooth(, tail = "both") and both tails are such that NA is returned, the current version gives a warning

In max(left_k, right_k, na.rm = TRUE) :
  no non-missing arguments to max; returning -Inf

as max() is not able to compare NA and NA.

This PR adds a check for left_k and right_k being both NA. The new test fails without the fix and passes with the fix.

This warning came up when using brms::loo_epred() which calls E_loo() which checks both tails, and new loo v2.10.0 has switched to using posterior package for the checks

I used Claude to find the error and suggest a fix. I understand the fix. I wrote the new test.

[github-actions]

This is how benchmark results would change (along with a 95% confidence interval in relative change) if f32efd9 is merged into master:

• ✔️as_draws_array: 141ms -> 142ms [-0.45%, +1.02%]
• ✔️as_draws_df: 68.6ms -> 68.5ms [-0.93%, +0.6%]
• ✔️as_draws_list: 155ms -> 154ms [-0.77%, +0.72%]
• ✔️as_draws_matrix: 24.5ms -> 24.5ms [-0.81%, +0.64%]
• ✔️as_draws_rvars: 113ms -> 113ms [-1.27%, +0.14%]
• ✔️summarise_draws_100_variables: 731ms -> 730ms [-0.32%, +0.07%]
• 🚀summarise_draws_10_variables: 82ms -> 81.8ms [-0.56%, -0.05%]
  Further explanation regarding interpretation and methodology can be found in the documentation.