wheels-build: do not fail fast

[jameslamb]

Proposes allowing all wheels-build jobs to run to completion, instead of immediately cancelling all of them when 1 fails.

This should help with narrowing down issues, like "only fails on arm64".

Motivation

Noticed while working on rapidsai/cugraph#5479, where @ChuckHastings and I were investigating an issue that only occurred on CUDA 12 builds. Noticed that all conda-cpp-build jobs ran to completion, but wheels-build jobs were cancelled after the first failure.

Notes for Reviewers

This change will increase CI runner usage, but I think it's worth it:

• CPU-only runners used for builds are generally cheap and plentiful
• failed wheel build jobs are valuable:
  • they provide information to speed up debugging (e.g. "only fails on arm64")
  • they partially fill up the shared build caches and package caches, which other CI jobs benefit from and which might reduce the time-to-resolution for issues

[jameslamb]

/merge