fix(metrics): bound histogram sample storage to prevent unbounded memory growth

[UGVicV]

Description

This PR resolves the issue where MetricsCollector.observe() stored every raw observation value in an unbounded list, causing unbounded memory growth in long-running agents with high-volume metrics.

Fix

• Replaced raw list storage with a new _HistogramBucket class using __slots__ for memory efficiency.
• Each bucket tracks: count, total (sum), min_val, max_val, and bounded samples (deque with maxlen).
• Added MAX_HISTOGRAM_SAMPLES = 1000 constant — only the most recent 1000 samples are retained.
• Aggregates (count, sum, min, max) are computed incrementally O(1) per observation, not recomputed from raw data.
• Snapshot now includes min and max fields alongside existing count, sum, avg.
• Wrapped observe() in try-except with logger.error() for robust error logging.
• Replaced print() with logger.error() in snapshot().
• Removed unused List import.
• Ensured 100% PEP8 compliance (all lines under 79 characters).

Verification (Proof)

Added 4 regression tests to tests/test_metrics.py:

• test_histogram_min_max: Verifies min/max tracking across observations.
• test_histogram_bounded_samples: Verifies memory is bounded after 1500 observations while count/sum remain accurate.
• test_histogram_aggregates_accurate: Verifies count/sum/avg/min/max for a known set.
• test_histogram_empty: Verifies no crash when querying unobserved metrics.

Test Execution Output

69 passed in 0.07s

Lint & Formatting Output

• flake8 src/common/metrics.py tests/test_metrics.py -> Passed (0 errors/warnings)
• git diff --check -> Passed (0 errors)

No secrets, tokens, or hidden context are included in the code.

Closes #982.