[NitinKumar004]

This lays the foundation for the ElastiCache parity work in #1472. It doesn't add new operations. It fixes how errors reach the wire and makes the control plane safe under concurrent use.

Typed faults. Provider errors now carry their AWS fault code in a typed field (Fault, with AWSFault()). Before, the code was a prefix in the message. The wire layer takes the code from that field, picks the HTTP status from a table that follows the botocore model, and writes the message with no code prefix. This replaces the old substring matching, which had two bugs:

• a duplicate parameter group came back as CacheClusterAlreadyExists; it is now CacheParameterGroupAlreadyExists;
• CacheSubnetGroupNotFoundFault came back as 404; it is now 400, as in AWS.

Control-plane lock. Mock gets a mutex, mu, that guards the cluster, replication group, subnet group, parameter group and snapshot stores, and the tag map. The separate tag mutex is gone: every tag access already ran under mu. The lock order is documented on the struct: mu, then a settle set's lock. Settle sets have no callbacks. No lock is held across a call to monitoring, to the VPC subnet resolver or to a real cache engine.

• A replication group that is still being created owns its member ids, so a cluster named <group>-001 is refused during that window. Creating a group whose member id is already used by a cluster returns CacheClusterAlreadyExists.
• Creates are atomic, so concurrent creates of the same id produce exactly one winner.
• The subnet-group and parameter-group in-use checks run in the same critical section as the delete. A cluster or group that is still being created counts as a user.
• Persist Snapshot and Restore run under mu.
• Parameter-group overrides are copied on write, so a Describe result handed out earlier is never mutated.

Id reservation around real engines. A create reserves the id while the real engine is being provisioned. A delete keeps the id reserved until Deprovision returns, so the same engine slot is never provisioned twice.

• During those windows, Describe and List return the resource with status creating or deleting.
• Mutating calls in those windows return InvalidCacheClusterState or InvalidReplicationGroupState, and a create of the same id returns AlreadyExists.
• If a teardown fails, the resource is put back.

This also locks the remaining parameter-group operations and the snapshot operations, which covers #1690 and #1697. Each has its own race test. Tagging snapshot ARNs is still not supported and stays not-found (#1695). It adds docs/coverage/nongoals/cache.md, which also notes two behaviours: a persisted snapshot leaves out resources that are still creating or deleting, and a deleting group's member clusters drop out of DescribeCacheClusters first.

Tests:

• 50-goroutine -race tests for duplicate creates of clusters, replication groups, subnet groups and parameter groups.
• TestModifyVsResetParameterGroupRace (AWS ElastiCache: lock the remaining parameter-group ops #1690) and TestSnapshotVsDeleteReplicationGroupRace, which runs snapshot create, copy and delete against deleting the group (AWS ElastiCache: lock the snapshot ops #1697).
• TestCreateCacheCollidingWithReservedGroupMember.
• Delete-vs-create races for subnet groups and parameter groups.
• Engine-gated tests: describing during creating and during deleting, mutating calls getting the invalid-state fault, recreate blocked until Deprovision returns, and Snapshot/Restore running alongside creates.
• TestWriteErrStripsCodeFromMessage, TestWriteErrStatusFromTable and an SDK test that checks the typed fault types, the status codes and prefix-free messages.

Verified:

• go build ./... passes.
• -race -count=3 passes on the ElastiCache provider and server packages, plus TestWiringParity, persist and TestAWSErrorMessageShapeCompat.
• golangci-lint on the touched packages shows no new issues against the development baseline.
• go generate ./... is clean after the commit.
• AWS CLI against cloudemu serve --async-settle: create-cache-cluster and create-replication-group report creating; duplicates return CacheClusterAlreadyExists and ReplicationGroupAlreadyExists with messages that carry no prefix; a missing subnet group returns HTTP 400; delete then recreate works.
• Terraform (hashicorp/aws 6.70.0) with aws_elasticache_parameter_group, aws_elasticache_cluster and aws_elasticache_replication_group: apply, plan shows no changes, update (parameter value and tags), plan shows no changes, destroy.

Part of #1472