refactor: abstract object storage backend

[dylayed]

Summary

This refactors Celld's S3 storage code so the same runtime paths can support another object-store backend. Storage setup is currently spread across bucket access, fleet startup, and LTX replication. This patch moves that setup behind a private storage_backend module and passes configured object_store::ObjectStore instances to callers.

This patch is S3-only and is intended to preserve existing S3 behavior.

Approach

• Keep bucket, credentials, retry, and client setup in Celld, where daemon configuration already lives.
• Use the existing object_store::ObjectStore trait instead of adding another provider abstraction.
• Treat the conditional-write identifier as an opaque version string. S3 still maps it to an ETag; the next patch can map GCS to a generation without changing ownership logic.
• Inject the configured store into celld-ltx instead of moving Celld's provider policy into that crate.

Most of the remaining diff is mechanical propagation through callers.

Review

The main things to review are the boundary in crates/celld/storage_backend.rs and whether existing S3 construction, credentials, retries, endpoint behavior, CAS behavior, and LTX configuration remain unchanged.

Review this PR before #1. The two PRs are linked in native stack #3.

Validation

• Celld library and test checks pass.
• The focused S3 object-storage tests pass.
• The four-scenario CAS contract passes against real MinIO.
• The existing MinIO LTX suite passes all 3 tests, including multipart.
• The PR refactor: abstract object storage backend #2 range passes git diff --check.