Improve training telemetry guards

[vikranthreddimasu]

Summary

• Validate throughput loss and timestamp metrics before they enter tracker history.
• Require loss recording inside a step context and only once per step.
• Fix dataset-size remediation so tiny datasets never suggest impossible batch_size=0.
• Keep optional YAML config import mypy-clean without making PyYAML mandatory.

Security changes

• No authentication, authorization, credential, or secret-handling behavior changed.
• Diff-only secret scan found no matches.

Usability changes

• Training preflight errors now give valid remediation for impossible min_batches cases.
• Telemetry failures now fail fast with direct ValueError/RuntimeError messages instead of poisoning dashboard/scheduler state.

Tests/checks run

• python -m pytest tests/test_monitoring/test_throughput.py -q
• python -m pytest tests/test_training/test_guards.py -q
• make lint
• make test outside sandbox: 1010 passed, 3 skipped, 2 warnings
• git diff --check master...HEAD
• git diff master...HEAD -- macfleet/ tests/ | rg secret-pattern scan: no matches

Commit list

• e1cab1a Validate throughput loss metrics
• cf692ae Fix dataset guard remediation advice
• 49da166 Guard throughput loss recording lifecycle
• 3a42601 Keep optional YAML import mypy-clean

Risks or manual review notes

• Low risk: changes are validation/error-message focused in telemetry, training guards, and optional YAML typing.
• Initial sandboxed make test failed because localhost bind was blocked; rerun outside sandbox passed.

Whether the PR was merged

• Not merged yet; waiting for GitHub checks.