Rework session storage: overflow-driven idempotent segments

[lyndonbauto]

Summary

Replaces the hot-tail-on-session-record + sealed c: chunk model with append-only K_ORDERED segment records (app:user:session:g:NNNNNNNN). Each segment holds one events Map keyed "{ts_micros:020d}:{event_id}".

• Idempotent append. The map key is a pure function of the event, so a retried/duplicated map_put overwrites the same slot — a timed-out-then-retried append can never duplicate an event.
• Cheap reads. Key-ordered map → server-side last-N via map_get_by_index_range(-N, N) and after_timestamp via map_get_by_key_range.
• React, don't predict. The only rollover trigger is a real RecordTooBig; no client-side byte estimation/threshold/flush. cur advances with a cur == N guarded increment so concurrent rollovers converge on the same next index.
• 1-RTT hot path. No-state-delta append = one operate; an append carrying state = one batch_write coalescing the segment write with session/app/user state writes (multi-scope appends now match single-scope latency).
• Back-pressure resilience. Because every hot-path write is idempotent, transient back-pressure is retried with bounded jittered backoff: DeviceOverload always (rejected, never applied) and ambiguous TimeoutError where a re-apply is a no-op, including per-record DeviceOverload in the coalesced batch_write. create_session retries DeviceOverload only (POLICY_EXISTS_CREATE put is not timeout-idempotent).

Fixes

• Heavy-concurrency data loss / unrecoverable RecordTooBig on a single hot session (events were dropped when concurrent appends overran the 1 MiB write-block-size).
• A lone event exceeding max-record-size now raises a clear RecordTooBig instead of looping.

Breaking

• Session on-disk layout changed (clean break, no migrator). Sessions written by <= 0.0.2 are not read by this version. Version bumped to 0.1.0.
• Removed flush_threshold_bytes / huge_event_bytes / max_tail_bytes ctor params, codec.estimate_event_size, keys.chunk_key, and the tbytes / chunks / seq / ts_lo / ts_hi / cidx bins.

Also includes the previously-staged Atomic Session Append tutorial (updated for the segment model).

Test plan

• Unit suite (no Docker): pytest -m "not aerospike" → 53 passed
• Session integration suite (testcontainers): pytest tests/test_sessions.py → 43 passed
• New server-free retry unit tests: tests/test_session_retry.py → 5 passed (overload retried/exhausted, timeout retried/opted-out, other errors propagate)
• Hot-path benchmark vs patched old model: no regression (faster under concurrency; multi-scope appends up to ~3x faster via batch_write coalescing)