Retry DAP source acquisition with exponential backoff

What does this MR do and why?

Transient Gitaly pack-objects load shedding can fail source acquisition before a DAP agent starts. Adds a retry at the startup stage when git access happens without restarting agent execution or replaying its side effects.

Behind the existing default-off, project-scoped dap_session_source_fetch_retry flag, configure:

• GET_SOURCES_ATTEMPTS: 6 — six total source-acquisition attempts.

References

• Source-acquisition failures: https://gitlab.com/gitlab-org/gitlab/-/work_items/627944
• Feature flag rollout: #628054
• Related work: #627739 (closed)

Screenshots or screen recordings

Before	After

How to set up and validate locally

1. Insert the following diff in your local gitaly and restart it to simulate gitaly failure:

diff --git a/internal/gitaly/service/smarthttp/upload_pack.go b/internal/gitaly/service/smarthttp/upload_pack.go
index 2fdd6b0e5..5a3213bed 100644
--- a/internal/gitaly/service/smarthttp/upload_pack.go
+++ b/internal/gitaly/service/smarthttp/upload_pack.go
@@ -19,6 +19,9 @@ import (
 )

 func (s *server) PostUploadPackWithSidechannel(ctx context.Context, req *gitalypb.PostUploadPackWithSidechannelRequest) (*gitalypb.PostUploadPackWithSidechannelResponse, error) {
+       // TEMPORARY TEST INJECTION: always shed load to demo DAP get_sources retry
+       return nil, structerr.NewResourceExhausted("resource exhausted, please try again later")
+
        repoPath, gitConfig, err := s.validateUploadPackRequest(ctx, req)
        if err != nil {
                return nil, structerr.NewInvalidArgument("%w", err)

2. Try a DAP job (e.g. mention Duo Developer), see it fail after a single attempt
3. Turn on dap_session_job_retry for the project / instance
4. Try a DAP job again, see it retry 6 times

MR acceptance checklist

Evaluate this MR against the MR acceptance checklist. It helps you analyze changes to reduce risks in quality, performance, reliability, security, and maintainability.

• Bounded retries remain before agent execution; whole-job retry removed.
• Default-off project flag retained, with enabled/disabled and resume coverage.
• Focused specs, RuboCop, and whitespace checks passed.
• Verify Runner compatibility and fault-injection behavior before rollout.

Changelog note: this retry behavior is behind a default-off flag. When squashing, remove the first commit's Changelog: fixed trailer. No history rewrite is included here.