enhance: scope collection load and state reads

[sunby]

Collection-scoped load and state reads currently rebuild global config or shard snapshots after unrelated updates. This change limits those reads to the requested collections, so their in-memory work scales with the selected configs and shards.

• Add atomic point reads and scoped load-config snapshots, retaining both collection and store versions without refreshing the full cached snapshot.
• Resolve Balancer trigger scope before capturing configs. Full listing, full reconciliation, conservative fallback, and cluster-wide row totals retain their existing behavior.
• Use scoped reads for LoadCollection admission, ShowLoadCollections/ShowLoadPartitions, GetQueryViewLoadInfo, and the load-progress path used by GetLoadState. Preserve current-replica filtering and the RPC's global version semantics.
• Document the read contract and existing manager-reclamation behavior. Add regressions for concurrent updates, release/reload, missing IDs, replica filtering, manager removal, and scoped/full triggers.

Related issue: this fork has issues disabled, so an associated issue could not be created.

Validation

• Full loadmgr, coordview, and balancer package tests passed with -race -tags dynamic,test -gcflags='all=-N -l'.
• Focused QueryCoord tests passed with the same flags: scoped load/progress reads, runtime recovery/notifications, segment-load-info watch sessions, and load/release ACK callbacks.
• golangci-lint --new-from-rev=HEAD passed for all four changed packages, including the staged new tests. run_clang_format.sh and git diff --check passed.
• make lint-fix was run but is blocked by an existing base-branch typecheck error: internal/metastore/kv/querycoord/kv_catalog_test.go:372:15: undefined: mocks. Formatter edits outside this change were reverted.

In-memory benchmark

Median of three 300 ms runs on Go 1.26.5, darwin/arm64, Apple M5, with -tags dynamic,test -gcflags='all=-N -l'. Each iteration invalidates the full snapshot and reads one collection with one replica/shard. Persistence, RPCs, and concurrent contention are excluded.

Loaded collections/shards	Config full (ms)	Config point (ns)	Config scoped (ns)	Shard full + filter (ms)	Shard scoped + filter (ns)
10,000	1.229	48.13	363.4	0.508	236.8
50,000	6.817	47.91	330.0	2.609	245.2
150,000	37.669	48.38	317.1	12.203	240.4

Point reads allocate zero bytes; scoped config and shard reads allocate 608 B and 352 B respectively at every size. These results validate the scoped in-memory cost; end-to-end UAT latency improvement has not been measured or deployed.