Segment membership: Bitmap sharding and parallel backfill

[khvn26]

Together with Flagsmith/flagsmith-engine#299, this unlocks:

• read-latency SLO tightening for 100M+ environments
• backfill SLO bucket tightening for 10M+ environments

Scope

• Migrate AtomBitmap from OneToOneField to (atom, shard_id) unique. Background job to split existing blobs by ord range.
• Read path fetches all shards for an atom in one query and unions in memory.
• process_segment_canonical_changed becomes a fan-out task spawning one child task per shard.

Acceptance criteria

• Per-shard blob caps at ~128 KB at the 870M-identity scale.
• Read latency p95 enters the "≤ 10M" SLO envelope even for the 870M env after sharding.
• Backfill of the 870M env completes in less than 24h on N workers (N TBD by review).