Typed column metadata reads via RetrieveMetadataChanges (one request, no @odata.type cast round-trip)

Bắt đầu trong data/_odata.py và kiểm tra cast path hiện có của metadata, sau đó lần theo list_columns và get_column trong operations/tables.py cùng với async twin của chúng. Bổ sung unit coverage cho request shape, projection, criteria và typed mapping của RetrieveMetadataChanges, đồng thời cập nhật README và CHANGELOG. Hoàn tất khi typed reads trả về các thuộc tính như MaxLength hoặc MinValue/MaxValue trong một request, trong khi vẫn giữ cast path hiện có làm fallback.

Summary

Reading type-specific column properties over /EntityDefinitions is a two-step dance: first read the generic attribute to get AttributeType/@odata.type, then re-issue a second GET with a derived-type URL cast (/Attributes/Microsoft.Dynamics.CRM.StringAttributeMetadata) to project the typed fields. The SDK does exactly this today (e.g. _bulk_fetch_picklists, extra_select="@odata.type,AttributeType").

Context: what the .NET modern client gives for free

ServiceClient gets fully-typed, polymorphic AttributeMetadata derived objects back in one RetrieveEntity/RetrieveAllEntities call and the caller just downcasts (attr is StringAttributeMetadata s). No second request, no discriminator juggling.

Proposed change

Add a typed read path backed by the Web API RetrieveMetadataChanges function -- the closest Web-API analog to the .NET single-request typed retrieve. It returns derived typed attributes in one call and supports property projection (via the query's AttributeQuery.Properties), so callers get small, typed payloads without the cast round-trip.

cols = client.tables.list_columns("account", typed=True)   # typed values, one request
col  = client.tables.get_column("account", "emailaddress1", typed=True)

Why this does not need a platform change

Payload size is controlled by the query's property projection, so the platform "omit null fields" ask is not a blocker -- we simply project only the properties we need. (Note: the .NET client itself uses plain Retrieve* + a client cache rather than RetrieveMetadataChanges; for a Web-API client wanting small typed payloads, RetrieveMetadataChanges with projection is the better fit. Design discussion welcome.)

Touchpoints

• data/_odata.py -- new _retrieve_metadata_changes(...) issuing GET RetrieveMetadataChanges(Query=@p1)?@p1={...} with an EntityQueryExpression + AttributeQuery.Properties projection.
• operations/tables.py + async twin -- typed=True on list_columns / get_column (keep the existing cast path as fallback).
• Unit tests: request shape (projection, criteria) + typed mapping.
• README + CHANGELOG.

Acceptance

• list_columns(typed=True) returns typed values (e.g. MaxLength on string/memo, MinValue/MaxValue on int) in a single request, no @odata.type URL cast.
• Property projection limits the payload to requested properties.

Related

• #194 -- create-time column constraints (in progress).
• microsoft/Dataverse-skills#119 -- documents the current two-step typed-read workaround.