fix: compare_windows uses rate-of-change for counter metrics

[tallpsmith]

Summary

Fixes #26.

• pcp_compare_windows was comparing raw cumulative counter values across windows — counters are monotonic odometers, so comparing means is meaningless (a counter at 1M vs 2M looks like "2x increase" when the rate was identical)
• Now fetches metric semantics via /series/descs, detects counters, and computes rate-of-change before stats comparison
• Instant/discrete metrics keep existing raw-value behavior

Changes

• _compute_rates() in _stats.py — converts timestamped counter samples to per-second rates, clamps to zero on counter wraps, skips zero-dt pairs
• _fetch_descs() in _fetch.py — fetches metric semantics from /series/descs with graceful fallback to "instant" on error
• _compare_windows_impl() in comparison.py — builds semantics-by-key map, rate-converts counters, adds "semantics" field to output
• All existing tests unaffected (they mock /series/descs as [] → defaults to instant)

Test plan

• _compute_rates: empty, single, monotonic, counter wrap, zero-dt, steady rate
• _fetch_descs: happy path, empty input, pmproxy error fallback
• Counter-aware comparison: counter uses rates, instant uses raw, descs failure fallback
• All 471 tests pass, 89% coverage
• Pre-push sanity clean
• E2E not run (no PMPROXY_URL set)