fix: avoid empty streaming query shards

[sunby]

After an SN restarts before persisting a Ready view as Up, a delayed Up/Dropped push can register an empty shard. Vchannel-only query planning may then return VIEW_NOT_FOUND even though a replacement replica has an Up view.

Create missing shards only for batches containing Preparing; otherwise report the lost view without indexing a shard. For unknown replica IDs, try the other vchannel candidates when a candidate has no Up view. Preserve explicit-replica lookup, context errors, query leases, and draining of accepted reports during WAL handoff.

Validation:

• The full snview suite passed with -race -tags dynamic,test -gcflags="all=-N -l" -count=1 on the target base; nine related QueryPlan/adaptor/client tests and the dynamic package build also passed.
• Regression coverage includes Ready-state loss and replacement, mixed batches, fixed-order unavailable candidates, cancellation, lease release, and concurrent handoff. The old implementation and both fallback/drain mutations triggered the expected regression failures.
• Targeted golangci-lint reported zero issues; formatter checks, run_clang_format.sh with clang-format 15, and git diff --check passed.
• make lint-fix was run but stopped at the existing internal/metastore/kv/querycoord/kv_catalog_test.go:372: undefined: mocks typecheck error. Unrelated formatter changes were restored. No cluster deployment was performed.

Related: #16.

issue: milvus-io#40451