perf(abcip): native read window >= 64; cold waiters hold no window slot

P8: on the native transport the bounded-parallel read window is also the cap
on services per Multiple Service Packet (the device's shared scheduler packs
whatever is queued, within the connection's request/reply byte budget), so the
default 8 cost a 1000-tag controller ~125 round trips per poll instead of ~10.
The effective window (AbCipDriver.ReadWindow) is now max(64, option) when the
native tag factory serves reads; libplctag and injected factories keep the
option exactly. Option doc + AdminUI help text updated (Razor: Wave 4 live
check).

P7 addendum: a work item now takes its window slot only after its runtime
exists. A cold create queues on its device's single-flight connect gate, and
holding a slot there let one dead device's backoff-expiry attempt occupy the
whole window for a full Timeout, stalling every other device's reads.

Refs #737
Claude-Session: https://claude.ai/code/session_01Ub9hGmYLQvn9XETzw7CMsT

src/Server/ZB.MOM.WW.OtOpcUa.AdminUI/Components/Shared/Drivers/Forms/AbCipDriverForm.razor

@@ -21,7 +21,7 @@
             <div class="col-md-3">
                 <label class="form-label">Max concurrent reads</label>
                 <InputNumber @bind-Value="_form.MaxConcurrentReads" @bind-Value:after="EmitAsync" class="form-control form-control-sm" />
-                <div class="form-text">Read work items in flight at once. Default 8; 1 = sequential.</div>
+                <div class="form-text">Read work items in flight at once. Default 8; 1 = sequential (libplctag). The native transport always uses at least 64 — it packs concurrent reads into one request.</div>
             </div>
             <div class="col-md-3">
                 <div class="form-check form-switch mt-4">