fix(manifest): parse same-indent block sequences instead of dropping them (issue #77)

[dndungu]

Summary

POST /api/v1/pods with a Pod manifest whose containers: list uses the
"dash at the same indentation as its key" YAML style (very common,
including in hand-written K8s manifests copy-pasted from docs) silently
dropped the entire container list. The pod was admitted, podman pod create ran, but no container was ever started -- the pod reported
completed/exitCode=0 within seconds with nothing having actually run.
This fixes internal/manifest's hand-rolled YAML parser to recognize that
style.

Changes

• internal/manifest/yaml.go: parseYAMLLines's "key with no inline
  value" branch only treated the next line as a list when it was indented
  deeper than the key (nextIndent > baseIndent). A sequence at the
  same indentation as its key (nextIndent == baseIndent) fell through
  to the nextIndent <= baseIndent case, which means "key has no value" --
  silently setting the key to "" and, because the top-level parse loop
  then also saw the now-unconsumed - line as a dedent, truncating the
  rest of the document parse entirely. Added an explicit branch for
  nextIndent == baseIndent && strings.HasPrefix(nextTrimmed, "- ") ahead
  of the existing catch-all, routing it through the same parseYAMLList
  already used for the deeper-indented case. Purely additive: it only
  changes behavior for the previously-mishandled case, so it can't affect
  any manifest that parsed correctly before.
• internal/manifest/yaml_test.go: TestParseYAML_SameIndentBlockList
  (bare-scalar sequence) and TestParseYAML_SameIndentBlockList_Nested
  (map-valued items under a nested key, matching the issue's actual
  shape) at the ParseYAML unit level.
• internal/manifest/same_indent_list_test.go:
  TestIssue77_SameIndentContainersList, an end-to-end Parse()
  reproduction using the exact manifest text from the issue (the
  GitHub token/URL placeholders scrubbed; irrelevant to the parse path
  under test).

Root cause vs #73/#85

internal/executor/podman.go's buildRunArgs/GPU-env-injection family
(#73, #85) corrupts Command/Args after a successful parse. This bug
is upstream of that: spec.Containers was already empty by the time
CreatePod ran, so its per-container loop had nothing to iterate --
podman run/podman container create was never invoked for the
workload container, only podman pod create for the pod shell. Confirms
docs/plan.md's R2/T2.4 question that #77 does not share #73/#85's root
cause.

Every existing YAML test fixture in this package used the extra-indented
containers:\n - name: ... style, so the same-indent style had zero
coverage before this change -- same "silently defaults instead of
erroring" pattern previously seen in #43, #44, #52, #66, #74.

Testing

• go build ./... -- clean.
• go vet ./... -- clean.
• staticcheck ./... -- clean.
• go test ./... -race -timeout 120s -count=1 -- all 13 packages pass,
  no regressions (internal/manifest 1.9s, full suite ~37s dominated by
  internal/watcher).
• go test ./internal/manifest/... -run 'TestIssue77|TestParseYAML_SameIndent' -v
  -- confirmed red before the fix (containers=0) and green after
  (containers=1, restartPolicy=Always) on all three new tests.
• Not live-verified on the DGX in this PR -- that happens centrally after
  merge/release per this wave's operating procedure. The exact repro
  manifest from the issue is now covered by an automated test; DGX
  live-verify just needs the real manifest (with the real token/URL)
  submitted once the release carrying this fix is deployed.

Linked issues / tasks

• Fixes Pod completes instantly (exitCode=0) with zero container startup — container spec silently dropped #77 (Pod completes instantly (exitCode=0) with zero container startup — container spec silently dropped): root-caused and fixed as
  described above.
• Progresses T2.4/T2.5 in docs/plan.md's E2 epic (issue Pod completes instantly (exitCode=0) with zero container startup — container spec silently dropped #77 investigate
  and fix); E2's remaining Wave 1 items (Pod command args: long quoted scalar mangled — bash -c receives no argument #73, podman pod stop/rm on a GPU-attached pod hangs indefinitely, and the stuck child becomes a zombie under the spark PID #88) are separate tasks in
  this same wave.