Allow buildkit to use the pod's credentials when running on k8s

When submitting a job to a k8s pod running buildkit-rootless:

buildctl --addr 'kube-pod://buildkitd-0?namespace=buildkit&container=buildkitd'  build ...

Our setup does not allow developers to push to ECRs, it must be done via a 'service' (e.g. buildkit in this case). However buildkit will always try to use the user's local credentials, even when the pod its running in has sufficient permissions (the same setup works for kaniko):

=> [auth] sharing credentials for xxxxxxxxxx.dkr.ecr.us-east-1.amazonaws.com                                                        0.0s
------
 > exporting to image:
------
error: failed to solve: failed to push xxxxxxxxxx.dkr.ecr.us-east-1.amazonaws.com/repo:tag: unexpected status: 403 Forbidden

Docs seems to suggest that when using S3 for caching, it's able to use IAM instance profiles / env vars / config files:

AWS Authentication:

The simplest way is to use an IAM Instance profile. Others options are:

Any system using environment variables / config files supported by the AWS Go SDK. The configuration must be available for the buildkit daemon, not for the client.

Can something similar be done when pushing to ECR? Instead of a user's local credentials being 'shared'..

Similar issues: https://github.com/moby/buildkit/issues/3947, https://github.com/moby/buildkit/issues/3772

Start with the buildctl kube-pod invocation and the daemon-side authentication behavior described in the issue, comparing ECR pushes with the documented S3 AWS authentication options. Determine whether the buildkit daemon can use the pod's IAM credentials instead of sharing the client's credentials, and define completion as a successful ECR push using only the pod's permissions.