Adds training-subsystem observability to INFO compression and
integration tests exercising the full training + retirement lifecycle.

Metrics (additive to the last_duration_ms / last_sample_count fields
wired by #39):
  compression_training_state          idle/scanning/submitted/cooldown
  compression_training_scans_started  total scans initiated
  compression_training_successes      successful dict promotions
  compression_training_failures       aborts + ZSTD errors
  compression_training_cooldown_until_ms
  compression_training_current_db
  compression_training_dicts_retired  dicts that completed lifecycle (GC'd)
  compression_training_debug_frame_refs  active retiring dict frame_refs

Integration tests (tests/integration/compression-training.tcl):
  - First-training trigger fires and produces a usable dict
  - After training, new writes get compressed
  - Training does NOT fire below min keys
  - Training aborts + cooldown on ineligible values
  - Training scans across multiple databases
  - Dict retirement: compress 20 keys -> decompression mode ->
    read all -> dict GC'd (asserts frame_refs == 20, the regression
    guard for the double-count fix in #42)

The retirement test is FLUSHALL-free: 200 training keys are SET before
the dict exists (stay RAW, no refs); only 20 keys SET after the dict
is active get compressed.

Depends on #42 (install-path double-count fix), now merged to unstable.

The first-training test had a confusing assertion:
  assert {$dict_after != $dict_before || $dict_before > 0}
The '|| dict_before > 0' clause made it vacuously true whenever a
dict already existed — the opposite of the intent. It was leftover
from shared-server (external) handling, but the test is external:skip
and always runs on a fresh instance where dict_before == 0.

Replaced with a direct fresh-instance check (active_dict_id == 0
before, > 0 after). Also dropped an unused dict_before capture in the
'does NOT fire' test (only known_dicts is asserted there).