[ML] Make changelog validation robust to shallow-clone merge base

[edsavage]

Summary

The Validate changelog entries step (.buildkite/scripts/steps/validate-changelogs.sh) fetches the PR base with --depth=1 and then runs a three-dot diff:

git diff --name-only --diff-filter=ACM origin/<base>...HEAD -- 'docs/changelog/*.yaml'

When a PR branched from an older point of the base branch, the shallow (--depth=1) histories share no common ancestor, so git aborts with:

fatal: origin/main...HEAD: no merge base

(exit 128). The diff runs inside a command substitution under set -euo pipefail, so the script dies at that line — before its own DIFF_EXIT/"skipping" soft-handler could run. The handler was effectively dead code, and the step hard-failed.

This is latent for any sufficiently-diverged branch; it surfaced on #3078 (see build 2895).

Fix

• Guard the diff with if ! VAR=$(...), which suppresses set -e for that command (the previous $? inspection could never be reached).
• On failure, progressively deepen both histories (--deepen 50/250/1000) to recover a merge base and retry.
• Only if that still fails, soft-skip with a warning.

This matters now that the overall build status (buildkite/ml-cpp-pr-builds) is a required check on main: a CI clone-depth limitation must never hard-fail an otherwise valid PR. Valid changelogs are still fully validated whenever a merge base is available (the common case, and now the deepened case).

Test plan

• CI green on this PR (its own changelog check is skipped via the >build label, exercising the early-exit path).
• Re-run the failing Harden model validation: reject __setstate__ hooks and forbid inductor::_reinterpret_tensor #3078 build after it picks up this change and confirm Validate changelog entries passes (merge base recovered, docs/changelog/3078.yaml validated).

[elasticsearchmachine]

Pinging @elastic/ml-core (Team:ML)