perf(cosmos): share partition key range cache across clients per endpoint

[tvaron3]

Summary

Clients targeting the same Cosmos DB endpoint now share a single CollectionRoutingMap cache instead of each maintaining an independent copy. This eliminates N-1 redundant copies of partition key range data when N clients connect to the same account.

Problem

When PPCB is enabled, each CosmosClient eagerly loads the full partition key range routing map via create_pk_range_wrapper(). Each client creates its own SmartRoutingMapProvider with its own _collection_routing_map_by_item dict, resulting in N identical copies for N clients targeting the same endpoint.

Changes

• _routing/aio/routing_map_provider.py + _routing/routing_map_provider.py: Module-level _shared_routing_map_cache dict keyed by endpoint URL. PartitionKeyRangeCache.init points _collection_routing_map_by_item to the shared entry. Added clear_cache() method.
• aio/_cosmos_client_connection_async.py + _cosmos_client_connection.py: refresh_routing_map_provider() calls clear_cache() instead of creating a new SmartRoutingMapProvider.

Memory Profiling Results

Test setup: tracemalloc, ~100 partitions, 2 regions, PPCB=True, 1 read + 1 upsert per client.

Current Memory (MB)

Clients	Original PPCB=True	Shared PPCB=True	Original PPCB=false	Shared PPCB=false
1	14.3	14.3	14.0	14.0
25	23.0	18.1	17.9	17.8
50	31.9	21.9	21.7	21.7
100	44.9	27.8	29.4	29.4
150	63.8	37.7	36.4	37.7

PPCB Overhead Reduction

Clients	Original	Shared Cache	Reduction
25	5.1 MB	0.3 MB	-93%
50	10.3 MB	0.3 MB	-97%
100	15.4 MB	-1.6 MB	-110%
150	27.4 MB	-0.0 MB	-100%

Scaling Estimate

At customer scale (200K partitions x 152 clients): ~2.1 GB -> ~14 MB (single shared copy).

Tests

5 unit tests in test_shared_pk_range_cache.py:

• Same endpoint shares cache (identity check)
• Different endpoints are isolated
• First client populates, second sees it
• clear_cache() resets the shared entry
• clear_cache() does not affect other endpoints

[tvaron3]

Moving to PR #46297 instead.