Ignore user-defined AWS_PROFILE variable in cache uploads

What does this MR do?

When the S3 RoleARN config is used, someone could define a CI/CD variable was AWS_PROFILE environment variable and cause the AWS SDK to look up the profile from config files. If the profile does not exist, the cache uploader fails with an error: failed to get shared config profile.

To avoid this, make sure we pass an empty AWS_PROFILE variable to the runner helper.

Why was this MR needed?

Runner cache uploads fail if AWS_PROFILE is defined in a job with RoleARN defined.

What's the best way to test this MR?

1. Set up a runner with a Docker executor and S3 cache.

2. Set up S3.RoleARN config: https://docs.gitlab.com/runner/configuration/advanced-configuration/#the-runnerscaches3-section

3. Create a CI job that defines AWS_PROFILE:

default:
  image: busybox:latest
  script:
    - echo "hello world" > test.txt
  cache:
    paths:
      - test.txt
  artifacts:
    paths:
      - test.txt
  variables:
    AWS_PROFILE: "random-profile"

What are the relevant issue numbers?

Relates to https://gitlab.com/gitlab-com/request-for-help/-/work_items/3883